An intelligent project success rate prediction method and device

Abstract

A project success rate intelligent prediction method and device, comprising: acquiring multi-source data, and constructing a structured knowledge base containing project success cases and their characteristics; applying the merged tree representation technology and the extended transformation to the project data in the structured knowledge base to construct a data structure, realize feature extraction and substructure search; based on the project feature set, multi-dimensional feature analysis is carried out from the time dimension and the space dimension; based on the multi-dimensional feature vector, a project-platform allocation model with group fairness is constructed; for the allocation model, a polynomial time approximation algorithm based on linear programming and network flow is designed to solve the optimal allocation scheme; according to the optimal allocation scheme and the historical project success data, the improved Reed-Solomon code technology is used for feature coding to construct a project success rate prediction model; based on the prediction model, the success rate of the project to be evaluated is evaluated and the support intensity trend is predicted. The present application improves the accuracy and efficiency of project success rate prediction.

Description

Technical Field

[0001] This invention relates to the field of data analysis and decision support, and in particular to a method and apparatus for intelligent prediction of project success rate based on multi-source data analysis. Background Technology

[0002] Project success rate prediction technology falls under the fields of data analysis and decision support. It aims to predict the probability of success for project applications, approvals, or implementation by analyzing historical data and relevant factors, providing decision-makers with a scientific basis. With the development of big data and artificial intelligence technologies, project success rate prediction has become an important tool for enterprises and institutions to improve resource utilization efficiency and reduce investment risks.

[0003] Traditional methods for predicting project success rates typically employ statistical regression analysis or decision trees. For example, some systems build simple predictive models by collecting basic data from historical projects, such as application time, applicant organization size, and industry. Other methods use expert experience scoring systems, where domain experts subjectively score various project indicators and estimate the probability of success accordingly.

[0004] Existing advanced prediction methods construct a knowledge base containing features of historically successful projects and combine it with machine learning algorithms for analysis. These methods typically extract features from a single dimension, such as considering only the technical characteristics of the project itself or the basic attributes of the applicant organization, and then use algorithms such as support vector machines or neural networks to build a predictive model to evaluate the success rate of new projects. The system analyzes patterns and trends in historical data, identifies key factors influencing project success, and generates prediction results accordingly.

[0005] However, existing technologies have significant shortcomings: First, most prediction methods analyze project characteristics from a single dimension, ignoring the comprehensive analysis of time and space dimensions, resulting in an inability to fully capture the multidimensional determinants of project success; second, existing methods typically employ hard constraints or simple threshold judgments, lacking flexible mechanisms for handling fairness among different project groups, and failing to balance overall utility with group fairness; finally, when dealing with large-scale datasets containing complex nested structures, traditional methods are computationally inefficient and difficult to adapt to real-time decision-making needs, especially when it is necessary to perform substructure search and pattern matching on semi-structured data such as JSON. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for intelligent prediction of project success rate, which realizes comprehensive project feature representation, fair resource allocation, and efficient prediction calculation, thereby improving the accuracy and efficiency of project success rate prediction.

[0007] To achieve the above objectives, this invention provides an intelligent prediction method for project success rate, comprising the following steps:

[0008] Acquire multi-source data related to the project, and construct a structured knowledge base containing successful project cases and their characteristics through data cleaning and standardization, named entity recognition, and relation extraction.

[0009] For the JSONL format project data in the structured knowledge base, the merging tree representation technique is applied to preserve the identity of individual projects while constructing a unified tree structure representation. The data structure is constructed through the extended Burrows-Wheeler transformation. Combined with path decomposition, ancestor calculation and adaptive tree identifier collection techniques, feature extraction and substructure search are realized to obtain the project feature set.

[0010] Based on the project feature set, the core dynamics of project development are analyzed from the time dimension, and the fundamental characteristics of the project enterprise are analyzed from the spatial dimension. Through feature fusion algorithm, a multi-dimensional feature vector is generated.

[0011] Based on the multidimensional feature vector, the project is divided into different groups, and a project-platform allocation model with group fairness is constructed. The overall platform constraints and group-specific constraints are modeled through a convex cost function to realize a soft constraint mechanism of restrictive dominance and minority group protection.

[0012] For the project-platform allocation model, a polynomial-time approximation algorithm based on linear programming and network flow is designed to solve for the optimal allocation scheme;

[0013] Based on the optimal allocation scheme and historical project success data, feature encoding is performed using improved folded Reed-Solomon code technology, and a project success rate prediction model is constructed using deterministic and randomized decoders.

[0014] Based on the project success rate prediction model, the success rate of the projects to be evaluated is assessed and the trend of support intensity is predicted, generating a visualized prediction report.

[0015] Preferably, the acquisition of multi-source data related to the project, through data cleaning and standardization, named entity recognition, and relation extraction, constructs a structured knowledge base containing successful project cases and their characteristics, including:

[0016] Based on web crawling technology and API interfaces, we collect multi-source data including lists of successful projects, guide documents, interpretation documents, and related news. We use regular expressions and text segmentation technology to perform preliminary processing on the collected unstructured data, remove irrelevant information, standardize the data format, and output the preliminary processed dataset.

[0017] The dataset that has been initially processed is cleaned using natural language processing techniques, including removing duplicate content, correcting typos, and standardizing the format. At the same time, named entity recognition technology is used to extract key entity information and output a standardized entity dataset.

[0018] Based on the standardized entity dataset, a relation extraction algorithm is applied to identify the relationships between entities, and knowledge graph technology is used to represent entities and relationships in a structured manner to construct a knowledge graph of successful project cases and output an initial knowledge base.

[0019] The initial knowledge base is optimized using knowledge reasoning and redundancy elimination algorithms. Distributed storage technology and indexing mechanisms are employed to store the optimized initial knowledge base as structured data in JSONL format, and the structured knowledge base is then output.

[0020] Preferably, the JSONL format project data in the structured knowledge base is processed by applying a merging tree representation technique to preserve the identity of individual projects while constructing a unified tree structure representation. A data structure is then built using an extended Burrows-Wheeler transformation. Combined with path decomposition, ancestor calculation, and adaptive tree identifier collection techniques, feature extraction and substructure search are achieved to obtain a project feature set, including:

[0021] For the JSONL format project data in the structured knowledge base, a JSON parser is applied to parse each project data into an independent tree structure, preserving the hierarchical relationship of key-value pairs, assigning a unique identifier to each tree node, ensuring that the original project can be traced in subsequent processing, and outputting a set of project tree structures.

[0022] Based on the aforementioned set of project tree structures, a tree merging algorithm is applied to merge the tree structures of multiple projects into a unified merged tree representation. During the merging process, the independent identity information of each project is preserved, and the same paths of different projects are distinguished by introducing identifier nodes and path prefixes, and the merged tree representation is output.

[0023] For the merged tree representation, the extended Burrows-Wheeler transformation algorithm is applied to convert the tree structure into an index data structure containing three parts: a node label array, a tree topology array, and an item identifier array. This supports fast tree navigation and sub-path querying, and outputs an extended Burrows-Wheeler transformed index structure.

[0024] Based on the extended Burrows-Wheeler transform index structure, combined with path decomposition, ancestor calculation and adaptive tree identifier collection technology, efficient substructure search of project data is achieved. Key features including technical field, innovation points, team size and funding needs are extracted from project data through predefined feature patterns, and a project feature set is output.

[0025] Preferably, the step of analyzing the core dynamics of project development from a time dimension and analyzing the fundamental characteristics of the project enterprise from a spatial dimension based on the project feature set, and generating a multi-dimensional feature vector through a feature fusion algorithm, includes:

[0026] Based on the project feature set, time attributes including project application time, approval time, and execution cycle are extracted. Time series analysis algorithms are applied to calculate the time distribution, seasonal fluctuations, and long-term trends of the number of projects, as well as the time patterns of changes in regulatory policies, and output time dimension feature vectors.

[0027] Based on the project feature set, spatial attributes such as geographical distribution, industry distribution, and scale distribution of the enterprises to which the projects belong are extracted. Spatial clustering and distribution analysis algorithms are applied to calculate the differences in project success rates in different regions and industries, as well as the correlation between enterprise scale and project success, and output spatial dimension feature vectors.

[0028] Based on the project feature set, we extract strength indicators including the company's asset size, R&D investment, team composition, and historical performance. We apply feature engineering techniques to calculate the comprehensive score and hierarchical classification of the company's strength, analyze the correspondence between the company's strength and project success, and output the company's strength feature vector.

[0029] For the time-dimensional feature vector, spatial-dimensional feature vector, and enterprise strength feature vector, a feature fusion algorithm is applied to weight and fuse the features of the time-dimensional, spatial-dimensional, and enterprise strength dimensions, taking into account the correlation and importance between features of different dimensions, to generate the multi-dimensional feature representation.

[0030] Preferably, the step of dividing projects into different groups based on the multidimensional feature vectors and constructing a project-platform allocation model with group fairness, and modeling the overall platform constraints and group-specific constraints through a convex cost function to achieve a soft constraint mechanism of restrictive dominance and minority group protection, includes:

[0031] Based on the multidimensional feature vector, a clustering algorithm is applied to divide the projects into different groups according to dimensions including industry category, technology field, and enterprise size. A weight coefficient is assigned to each group, and the project group division results are output.

[0032] Based on historical platform statistics and fair allocation standards, the service capacity limit of each platform and the acceptance ratio requirements of each group are determined. Convex function modeling technology is applied to construct a convex cost function of the overall platform constraint, so that when the total number of projects allocated to the platform exceeds the soft constraint, a non-linearly increasing cost penalty will be generated, and the platform constraint model is output.

[0033] Based on the project group division results, fairness constraint modeling technology is applied to construct a specific convex cost function for each platform and each group. This function promotes group fairness by penalizing unbalanced allocation, realizes a soft constraint mechanism of restrictive dominance and minority group protection, and outputs a group fairness constraint model.

[0034] Based on the matching degree assessment of projects and platforms, a utility value is defined for each project-platform pair, an overall utility function is constructed, and a comprehensive optimization objective function is constructed by combining the platform constraint model and the group fairness constraint model.

[0035] Preferably, the step of designing a polynomial-time approximation algorithm based on linear programming and network flow to solve for the optimal allocation scheme in the project-platform allocation model includes:

[0036] Based on the project-platform allocation model, a linear programming relaxation problem is constructed. The convex cost function in the original problem is transformed into a set of linear constraints through piecewise linearization, while retaining the utility constraints. This forms a relaxation problem that can be solved by standard linear programming methods, and outputs a linear programming model.

[0037] The linear programming model is solved using the simplex method or the interior point method. Fractional solutions are obtained during the solution process, and the optimal solution of the linear programming problem is output.

[0038] Based on the optimal solution of the linear programming, a corresponding network flow model is constructed, with the project as the source and the platform as the sink. The score allocation in the optimal solution of the linear programming is converted into the flow in the network. The minimum cost flow problem that satisfies capacity constraints and flow conservation is constructed, and the network flow model is output.

[0039] For the network flow model, the minimum cost flow algorithm is applied to obtain the integer flow solution, thereby determining the integer allocation scheme from the project to the platform. Theoretical analysis ensures that the cost difference between the approximate solution and the global optimal solution does not exceed the preset approximation ratio, while satisfying the utility constraint, and outputs the optimal allocation scheme.

[0040] Preferably, the step of using improved folded Reed-Solomon code technology for feature encoding based on the optimal allocation scheme and historical project success data, and constructing a project success rate prediction model through deterministic and randomized decoders, includes:

[0041] Based on the optimal allocation scheme and historical project success data, a model training dataset is constructed. Through data cleaning and standardization, including feature standardization, missing value handling, and outlier detection, the processed training dataset is output.

[0042] The processed training dataset is then used to perform feature encoding using an improved folded Reed-Solomon code technique, which maps the high-dimensional feature space to a low-dimensional encoding space and outputs the encoded feature data.

[0043] Preferably, the step of using improved folded Reed-Solomon code technology for feature encoding based on the optimal allocation scheme and historical project success data, and constructing a project success rate prediction model through deterministic and randomized decoders, includes:

[0044] Based on the encoded feature data, a deterministic decoder with near-linear time complexity is designed. The deterministic decoder can complete the decoding process in near-linear time and output a deterministic prediction model.

[0045] Based on the encoded feature data, a randomized decoder with a full polynomial time complexity is designed. The randomized decoder introduces a random sampling strategy to further reduce computational complexity while ensuring accuracy, and outputs a randomized prediction model.

[0046] Based on the deterministic prediction model and the randomized prediction model, a comprehensive prediction model is constructed using model ensemble techniques. The model parameters are optimized through cross-validation and grid search methods, and the project success rate prediction model is output.

[0047] Preferably, the step of assessing the success rate of the project based on the project success rate prediction model and predicting the trend of support intensity for the project to be evaluated, and generating a visualized prediction report, includes:

[0048] Based on the project success rate prediction model and the feature vector of the project to be evaluated, the success rate of the project is calculated under different platforms and conditions. The deterministic decoder or the randomized decoder is applied, and the appropriate decoding method is selected according to the real-time requirements of the application scenario to output the project success rate prediction result.

[0049] Based on the project success rate prediction results and historical support policy data, time series forecasting technology is applied to analyze the future trend of support intensity. Taking into account factors including policy cycles, economic environment and industry development, the support intensity of different types of projects is predicted, and the trend analysis results of support intensity are output.

[0050] Based on the project success rate prediction results and the support intensity trend analysis results, data visualization technology is used to generate an intuitive prediction report. The prediction report includes the success rate value, confidence interval, key influencing factor analysis, and support intensity trend chart, and the visualized prediction report is output.

[0051] The present invention also provides an intelligent prediction device for project success rate, comprising:

[0052] The data acquisition module is used to acquire multi-source data related to the project. Through data cleaning and standardization, named entity recognition and relation extraction, a structured knowledge base containing successful project cases and their characteristics is constructed.

[0053] The feature extraction module is used to extract features from the JSONL format project data in the structured knowledge base by applying the merging tree representation technique to preserve the identity of individual projects while constructing a unified tree structure representation. It also constructs a data structure through an extended Burrows-Wheeler transformation and combines path decomposition, ancestor calculation and adaptive tree identifier collection techniques to achieve feature extraction and substructure search, thereby obtaining the project feature set.

[0054] The multidimensional analysis module is used to analyze the core dynamics of project development from the time dimension and the fundamental characteristics of the project enterprise from the spatial dimension based on the project feature set, and to generate multidimensional feature vectors through feature fusion algorithms.

[0055] The allocation model construction module is used to divide projects into different groups based on the multidimensional feature vector, construct a project-platform allocation model with group fairness, and model the overall platform constraints and group-specific constraints through a convex cost function to realize a soft constraint mechanism of restrictive dominance and minority group protection.

[0056] The optimization solution module is used to design a polynomial-time approximation algorithm based on linear programming and network flow to solve for the optimal allocation scheme in the project-platform allocation model.

[0057] The prediction model building module is used to perform feature encoding using improved folded Reed-Solomon code technology based on the optimal allocation scheme and historical project success data, and to build a project success rate prediction model through deterministic and randomized decoders.

[0058] The results generation module is used to assess the success rate of the project and predict the trend of support intensity based on the project success rate prediction model, and generate a visual prediction report.

[0059] The beneficial effects of this invention are as follows:

[0060] This invention provides an intelligent prediction method and apparatus for project success rates. It innovatively merges multiple JSON object trees into a unified representation while preserving the identity of individual projects. It also incorporates an extended Burrows-Wheeler transform to construct an efficient index structure, enabling rapid substructure search and feature extraction of project data. By modeling platform-wide constraints and group-specific constraints using convex cost functions, it implements a soft constraint mechanism of restrictive dominance and minority group protection, balancing overall utility and group fairness. Combining linear programming relaxation and network flow techniques, it designs an efficient polynomial-time approximation algorithm to solve the project-platform allocation problem while maintaining the approximation ratio. It applies an improved folded Reed-Solomon code technique for feature encoding and designs a near-linear-time deterministic decoder and a polynomial-time randomized decoder, significantly improving the computational efficiency of project success rate prediction. Furthermore, it analyzes project data from multiple perspectives, including time and space, extracting features such as project quantity trends, changes in regulatory intensity, industry distribution, and geographical distribution. Multi-dimensional feature fusion is then performed to form a comprehensive project feature representation. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a flowchart of the intelligent prediction method for project success rate according to an embodiment of the present invention;

[0063] Figure 2 This is a flowchart of the spatiotemporal multidimensional feature analysis according to an embodiment of the present invention;

[0064] Figure 3 This is a structural block diagram of the intelligent prediction device for project success rate according to an embodiment of the present invention. Detailed Implementation

[0065] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0066] like Figure 1 As shown, the intelligent prediction method for project success rate in this embodiment of the invention includes the following steps:

[0067] Step S101: Obtain multi-source data related to the project, and construct a structured knowledge base containing successful project cases and their characteristics through data cleaning and standardization, named entity recognition and relation extraction.

[0068] This step begins by collecting multi-source data related to the project through various channels. Multi-source data refers to data sets from different sources with different formats and structures, including lists of successful projects, guide documents, interpretive documents, and related news. This data typically exists in unstructured or semi-structured form and requires a series of processing steps to transform it into usable structured knowledge.

[0069] Data cleaning refers to the process of identifying and correcting erroneous, inconsistent, redundant, or missing data in a dataset. In this invention, data cleaning includes operations such as removing duplicate content, correcting typos, and standardizing formatting to improve data quality. Standardization processing involves converting data from different sources and in different formats into a unified standard format to facilitate subsequent analysis and processing.

[0070] Named Entity Recognition (NER) is a fundamental technology in natural language processing used to identify and extract specific categories of entities from text, such as names of people, places, organizations, dates, and quantities. In this invention, NER technology is primarily used to extract key entity information such as project name, applicant organization, industry, and funding amount from project-related text.

[0071] Relation extraction is the process of identifying semantic relationships between entities, building upon the initial entity identification. For example, it can identify relationships such as "application-approval," "company-industry," and "project-funding amount." Through relation extraction, discrete entity information can be connected to form a structured knowledge network.

[0072] A structured knowledge base is a database that stores and organizes the extracted entities and relationships in a structured manner. In this invention, knowledge graph technology is used to represent and store this structured knowledge. A knowledge graph is a semantic network composed of nodes (entities) and edges (relationships), which can intuitively represent complex relationships between entities and support efficient querying and reasoning.

[0073] Step S102: For the JSONL format project data in the structured knowledge base, apply the merging tree representation technique to retain the identity of individual projects while constructing a unified tree structure representation. Construct the data structure through the extended Burrows-Wheeler transformation. Combine path decomposition, ancestor calculation and adaptive tree identifier collection techniques to realize feature extraction and substructure search, and obtain the project feature set.

[0074] This step processes the project data stored in the structured knowledge base in JSONL format. JSONL (JSON Lines) is a text format that facilitates the storage of structured data, where each line contains a single valid JSON value, typically an object or array. This format is particularly suitable for representing complex attributes of project success stories.

[0075] A key innovation of this invention is the tree merging technique, which allows multiple independent project tree structures to be combined into a unified tree while preserving the unique identity information of each project. During the merging process, the introduction of identifier nodes and path prefixes distinguishes identical paths across different projects. This representation method significantly improves data processing efficiency while maintaining the integrity and traceability of the original data.

[0076] The Extended Burrows-Wheeler Transform (xBWT) is an extension of the traditional Burrows-Wheeler Transform for indexing and compressing tree structures. While the traditional BWT is a reversible transformation algorithm for text compression and indexing, xBWT extends it to tree structures. In this invention, xBWT transforms the merged tree representation into three main arrays: a node label array (storing the labels of all nodes in the tree), a tree topology array (encoding the structural information of the tree), and an item identifier array (recording the item identifiers associated with each node). This indexing structure supports efficient tree navigation and subpath query operations.

[0077] Path decomposition is a technique that breaks down complex paths in a tree into a series of simpler path segments, facilitating rapid location and matching. Ancestor calculation is the process of determining the ancestor-descendant relationship between any two nodes, which is crucial for understanding the hierarchical relationships in a tree structure. Adaptive tree identifier collection is an optimization method that automatically adjusts the identifier collection strategy based on query requirements, improving query efficiency.

[0078] By combining these technologies, key features such as technical field, innovation points, team size, and funding needs can be efficiently extracted from project data to form a project feature set. This method avoids the tedious tree traversal process of traditional methods, significantly improving the efficiency of feature extraction.

[0079] Step S103: Based on the project feature set, analyze the core dynamics of project development from the time dimension and analyze the fundamental characteristics of the project enterprise from the spatial dimension. Generate a multi-dimensional feature vector through feature fusion algorithm.

[0080] In this step, a multi-dimensional analysis of the project feature set is conducted, mainly from the two dimensions of time and space, in order to obtain a more comprehensive representation of the project features.

[0081] Time-dimensional analysis focuses on the distribution and changing patterns of projects over time. First, it extracts fundamental time-related attributes, such as project application time, approval time, and execution cycle. Then, time series analysis algorithms are applied to calculate the temporal distribution, seasonal fluctuations, and long-term trends of project numbers. Time series analysis is a statistical method used to analyze data points arranged chronologically to identify patterns, trends, and cyclical changes. This analysis can reveal the temporal patterns of project applications and success rates, as well as the impact of regulatory policy changes on project success.

[0082] Spatial dimension analysis focuses on the spatial distribution characteristics of projects across different regions, industries, and sizes. It extracts spatial attributes such as the geographical distribution, industry distribution, and size distribution of the companies to which the projects belong, and then applies spatial clustering and distribution analysis algorithms to calculate differences and correlations across different dimensions. Spatial clustering is a technique for grouping objects with similar spatial characteristics, helping to identify groups of projects with common features. Distribution analysis, on the other hand, uses statistical methods to reveal the distribution of features across different categories, helping to understand the impact of various factors on project success.

[0083] In addition to time and space dimensions, the analysis also considers the company's fundamental characteristics, such as asset size, R&D investment, team composition, and historical performance. These indicators directly reflect the comprehensive capabilities and resource endowments of the project applicant and have a significant impact on the project's success rate.

[0084] Feature fusion algorithms are methods that organically combine features from different dimensions to generate more predictive comprehensive feature representations. Feature fusion considers the correlation and importance between features of different dimensions, integrating multi-dimensional features into a unified feature vector through methods such as weighted fusion, principal component analysis, or deep learning. This multi-dimensional feature representation can more comprehensively characterize various aspects of a project, providing a richer information foundation for subsequent allocation and prediction models.

[0085] Step S104: Based on the multidimensional feature vector, the project is divided into different groups, and a project-platform allocation model with group fairness is constructed. The overall platform constraint and group-specific constraint are modeled by the convex cost function to realize a soft constraint mechanism of restrictive dominance and minority group protection.

[0086] In this step, projects are first divided into different groups based on multidimensional feature vectors. This grouping is fundamental to achieving equitable resource allocation, as different types of projects may require different evaluation criteria and support policies. Clustering algorithms are used to divide projects according to dimensions such as industry category, technology field, and company size, and weight coefficients are assigned to each group to reflect its importance and priority within the overall project portfolio.

[0087] Group equity refers to the principle of ensuring equitable treatment of different groups in resource allocation. This invention primarily considers two core equity concepts: Restrictive Dominance and Minority Protection. Restrictive Dominance ensures that no group over-consumes resources, setting an upper limit for resource allocation; Minority Protection ensures that vulnerable or minority groups receive sufficient resource support, setting a lower limit for resource allocation. These two concepts together constitute the framework of group equity.

[0088] A convex cost function is a special mathematical function whose second derivative is always positive, and whose graph exhibits an upward convex shape. In this invention, the convex cost function is used to model platform-wide constraints and group-specific constraints. For platform-wide constraints, when the number of projects allocated to the platform exceeds its soft constraint (baseline capacity), a non-linearly increasing cost penalty will occur; for group-specific constraints, when the allocation ratio of a certain group on a certain platform deviates from the ideal quota, a corresponding cost penalty will also occur.

[0089] Soft constraints are a more flexible form of constraint compared to hard constraints. Hard constraints require strict adherence to conditions, while soft constraints allow for some degree of violation, but impose corresponding penalties. In this invention, a soft constraint mechanism is implemented through a convex cost function, enabling appropriate concessions in special circumstances while ensuring overall fairness through a cost mechanism. This method is more flexible than traditional hard quotas and achieves a better balance between efficiency and fairness.

[0090] Through the above design, a comprehensive project-platform allocation model was constructed. This model considers multiple factors such as platform capacity, group fairness, and allocation utility, providing a theoretical basis for subsequent optimization solutions.

[0091] Step S105: For the project-platform allocation model, design a polynomial-time approximation algorithm based on linear programming and network flow to solve for the optimal allocation scheme.

[0092] In this step, the project-platform allocation model constructed in the previous step is solved to obtain the optimal or near-optimal allocation scheme. Since the original model contains convex cost functions and integer constraints, direct solution may have high computational complexity. Therefore, this invention designs a polynomial-time approximation algorithm based on linear programming and network flow.

[0093] Linear programming (LP) is an optimization technique used to maximize or minimize a linear objective function under linear constraints. In this invention, the original problem is first transformed into a linear programming relaxation problem by relaxing integer constraints and linearizing the convex cost function. Integer constraint relaxation allows variables that were originally required to take integer values ​​to take fractional values; for example, relaxing a binary variable x∈{0,1} to 0≤x≤1. Convex cost function linearization approximates the original nonlinear convex function by using piecewise linear functions, such as introducing auxiliary variables and linear constraints to represent quadratic or exponential functions.

[0094] For the transformed linear programming problem, mature solution algorithms can be applied, such as the Simplex Method or the Interior Point Method. The Simplex Method is an iterative algorithm that finds the optimal solution by moving between the vertices of the feasible region; the Interior Point Method starts from the interior of the feasible region and approaches the optimal solution along the central path. These algorithms can solve large-scale linear programming problems in polynomial time and obtain the optimal solution for relaxed linear programming problems.

[0095] Network flow is a special type of linear programming problem that studies the allocation and optimization of traffic in a network. In this invention, the optimal solution of a linear programming problem is mapped to a network flow problem, where items are the source nodes, platforms are the sink nodes, and the fractional allocation in the linear programming solution corresponds to the traffic in the network. This mapping utilizes the special structure and properties of the network flow problem, especially its all-integer property: when all capacities in the network are integers, there exists an integer optimal flow.

[0096] The minimum cost flow problem is a type of network flow problem, aiming to find the flow allocation that satisfies the flow requirements while minimizing the total cost. Minimum cost flow algorithms, such as augmenting path methods, cost scaling methods, or network simplex methods, are applied to the constructed network flow model to obtain integer flow solutions, thereby determining the integer allocation scheme from the project to the platform.

[0097] Approximation algorithms are a class of algorithms that may not find the exact optimal solution, but can guarantee that the quality of the solution is within theoretical limits. In this invention, theoretical analysis proves that the designed algorithm can find a near-global optimal solution in polynomial time, with the cost difference not exceeding a preset approximation ratio. This algorithm design balances computational efficiency and solution quality, making it suitable for handling large-scale project-platform allocation problems.

[0098] Step S106: Based on the optimal allocation scheme and historical project success data, feature encoding is performed using the improved folded Reed-Solomon code technique, and a project success rate prediction model is constructed using deterministic and randomized decoders.

[0099] In this step, a project success rate prediction model is built based on the optimal allocation scheme and historical project success data. The core innovation of this step lies in applying an improved folded Reed-Solomon code technique for feature encoding and designing an efficient decoder.

[0100] Folded Reed-Solomon Codes (FRS) are error-correcting codes evolved from classic Reed-Solomon codes. Originally used for error correction in communication and storage, Reed-Solomon codes possess powerful multi-error detection and correction capabilities. Folded Reed-Solomon codes reduce code length through a "folding" operation while maintaining strong error correction capabilities. In this invention, FRS codes are innovatively applied to the field of feature coding, mapping a high-dimensional feature space to a low-dimensional coding space while preserving key information.

[0101] Feature encoding is the process of transforming raw features into a form more suitable for machine learning algorithms. In this invention, FRS encoding treats the raw features as information to be encoded and maps them to codewords through a specific encoding function. This encoding method has advantages such as reducing dimensionality, enhancing noise resistance, preserving feature structure, and improving computational efficiency.

[0102] A deterministic decoder is an algorithm capable of completing the decoding process in near-linear time. Near-linear time refers to a time complexity close to O(n·polylog(n)), where n is the codeword length and polylog(n) represents a polynomial of log(n). The deterministic decoder designed in this invention combines techniques such as Fast Fourier Transform (FFT), sparse recovery, and algebraic geometry, significantly reducing the computational complexity of traditional decoders to O(n²) or higher. This efficient decoder is particularly suitable for scenarios requiring real-time response.

[0103] A randomized decoder is a decoding algorithm that runs in full polynomial time by introducing a random sampling strategy. Full polynomial time refers to a time complexity of O(n^2). c ), where c is a constant. The randomized decoder designed in this invention utilizes the ideas of the Monte Carlo method and the Las Vegas algorithm to obtain the correct result with a high probability through multiple random attempts. Although the theoretical complexity may be slightly higher than that of the deterministic decoder, in practice, the constant factor is smaller, and it has better parallelism and the ability to adapt to complex noise.

[0104] Model ensemble is a commonly used technique in machine learning to improve prediction performance by combining the prediction results of multiple models to obtain more accurate and stable predictions. In this invention, a deterministic decoder model and a randomized decoder model are integrated using methods such as stacking, bagging, or boosting. The model parameters are optimized through cross-validation and grid search to finally construct a comprehensive prediction model.

[0105] The above design can significantly improve computational efficiency while ensuring prediction accuracy, providing strong support for predicting the success rate of large-scale projects.

[0106] Step S107: Based on the project success rate prediction model, assess the success rate of the project to be evaluated and predict the trend of support intensity, and generate a visualized prediction report.

[0107] In this step, the previously built prediction model is applied to conduct a comprehensive analysis and prediction of the new project to be evaluated, and finally generate an intuitive and visual prediction report.

[0108] First, feature extraction is performed on the project to be evaluated. The same method as in step S102 is used to obtain the key features of the project, and multidimensional feature analysis is conducted. Then, based on the project success rate prediction model and the extracted feature vectors, the success rate of the project is calculated under different platforms and conditions. Depending on the real-time requirements of the application scenario, a deterministic decoder (suitable for scenarios requiring rapid response) or a randomized decoder (suitable for scenarios requiring high accuracy) can be selected for prediction.

[0109] Support intensity trend forecasting is the prediction of future changes in the strength of policy support. Based on project success rate forecasts and historical support policy data, time series forecasting techniques are applied to analyze future trends in support intensity. Time series forecasting is a method of predicting future values ​​using historical data; commonly used techniques include moving averages, exponential smoothing, and ARIMA models. The analysis considers various factors such as policy cycles, economic environment, and industry development to make scientific predictions about changes in support intensity for different types of projects.

[0110] Visualization technology is a method of converting data into graphical representations, which can intuitively display patterns, trends, and relationships within the data. In this invention, data visualization technology is applied to transform success rate prediction results and support intensity trend analysis results into intuitive charts and reports. The prediction report includes success rate figures, confidence intervals, analysis of key influencing factors, and support intensity trend charts, providing decision-makers with comprehensive information support.

[0111] Through this visual forecasting report, decision-makers can clearly understand the project's success prospects and influencing factors, thereby making more scientific and targeted decisions.

[0112] In this embodiment, step S101 includes the following sub-steps:

[0113] Step S201: Collect multi-source data, including lists of successful projects, guide documents, interpretation documents, and related news, based on web crawling technology and API interfaces. Perform preliminary processing on the collected unstructured data using regular expressions and text segmentation technology to remove irrelevant information, standardize the data format, and output the preliminary processed dataset.

[0114] This sub-step involves collecting raw project-related data through various channels and methods. Web crawling is an automated program capable of browsing the internet and extracting specific information. Web crawlers are used to collect lists of successful projects, policy guidelines, interpretive documents, and related news reports from official websites, industry portals, and news media. APIs (Application Programming Interfaces) are interfaces for data exchange between applications. By calling the APIs of various data platforms and services, more structured and professional project data can be obtained.

[0115] The raw data collected is usually unstructured, meaning it lacks a predefined data model or organization. Preliminary processing is required to make this data usable for subsequent analysis. Regular expressions are used to match patterns in character combinations within a string; in this case, they are used to identify and extract specific patterns from the text, such as project numbers, application dates, and funding amounts. Text segmentation techniques are used to break down long texts into paragraphs, sentences, or semantic units, facilitating subsequent fine-grained processing.

[0116] During the initial processing, irrelevant information, such as advertising content, page navigation, footnotes, and other elements unrelated to the project analysis, will be removed. Simultaneously, data from different sources will be standardized, unifying data formats, field names, and value ranges to lay the foundation for subsequent data fusion and analysis.

[0117] This sub-step outputs a pre-processed dataset, which, while not yet fully structured and standardized, has removed most of the noise and irrelevant information and has a relatively uniform format.

[0118] Step S202: Apply natural language processing technology to clean the dataset after preliminary processing, including removing duplicate content, correcting typos and standardizing the format. At the same time, use named entity recognition technology to extract key entity information and output a standardized entity dataset.

[0119] In this sub-step, the initially processed dataset undergoes further cleaning and normalization. Natural Language Processing (NLP) is a field of computer science and artificial intelligence that studies the interaction between computers and human language. In this work, NLP techniques are used to understand and process text data, identifying its semantic information.

[0120] Data cleaning is a crucial step in data preprocessing, aiming to improve data quality. The first step is to detect and remove duplicate content to prevent the same information from being calculated multiple times. Duplicate detection can be based on methods such as exact matching, approximate matching, or semantic matching to ensure the uniqueness and representativeness of the data.

[0121] Typo correction is a crucial step in improving text quality. Spell checking algorithms, language models, or rule-based methods are used to identify and correct spelling and grammatical errors in text. This process not only improves readability but also lays the foundation for subsequent entity recognition and relation extraction.

[0122] Standardizing data formats is a necessary step to ensure data consistency. Normalizing data from different sources and in different formats, such as standardizing date formats, numerical representations, and units of measurement, enables effective comparison and calculation of the data.

[0123] Named Entity Recognition (NER) is the task of automatically identifying named entities in text and classifying them into predefined categories. In project analysis, important entity categories include project name, applicant organization, industry, funding amount, application date, and approval status. NER can be implemented based on rules, statistical models, or deep learning methods. Rule-based NER uses predefined patterns and dictionaries; statistical models such as Conditional Random Fields (CRF) utilize feature engineering and probabilistic models; deep learning methods such as BERT or BiLSTM-CRF can automatically learn feature representations and perform better on complex text.

[0124] This sub-step outputs a standardized entity dataset containing structured entity information extracted from the original text, such as projects, organizations, industries, and amounts. This information will serve as the foundation for building a knowledge graph.

[0125] Step S203: Based on the standardized entity dataset, apply a relation extraction algorithm to identify the relationships between entities, use knowledge graph technology to represent entities and relationships in a structured way, construct a knowledge graph of successful project cases, and output the initial knowledge base.

[0126] In this sub-step, based on entity identification, the semantic relationships between entities are further mined, and the entities and relationships are organized into a knowledge graph. Relation extraction is the process of identifying semantic relationships between specific types of entities from text. In project analysis, important relationship types include "application-approval," "company-industry," "project-funding amount," and "company-region," etc.

[0127] Relation extraction can be achieved through various methods. Rule-based methods use predefined grammatical and semantic patterns to identify relations; feature-based methods extract linguistic features from text and then use machine learning classifiers to predict relation types; neural network-based methods, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or graph neural networks (GNNs), can automatically learn deep feature representations of text, improving the accuracy and generalization ability of relation extraction.

[0128] A knowledge graph is a semantic network composed of nodes (entities) and edges (relationships), capable of representing complex networks of relationships between entities. In a knowledge graph, each node represents an entity (such as a project, enterprise, or industry), and each edge represents a semantic relationship between two entities (such as applying, belonging, or obtaining). The advantage of knowledge graphs lies in their ability to intuitively represent and store complex knowledge structures, support efficient querying and reasoning, and provide powerful support for subsequent data analysis.

[0129] The process of building a knowledge graph includes three main steps: entity linking, relation mapping, and graph construction. Entity linking matches entities identified in the text with existing entities in the knowledge base, solving the problems of entity disambiguation and coreference resolution; relation mapping maps extracted relations to predefined relation types or ontology models; and graph construction organizes entities and relations into a connected graph structure, forming a complete knowledge graph.

[0130] This sub-step outputs an initial knowledge base containing structured knowledge of project success cases, represented in the form of a knowledge graph. While this initial knowledge base contains rich entity and relationship information, it may still have redundancy, inconsistencies, or incompleteness, requiring further optimization.

[0131] Step S204: Optimize the initial knowledge base by applying knowledge reasoning and redundancy elimination algorithms, and use distributed storage technology and indexing mechanism to store the optimized initial knowledge base as structured data in JSONL format, and output the structured knowledge base.

[0132] In this sub-step, the initial knowledge base is optimized and standardized to improve the quality and usability of the knowledge, and an efficient storage and indexing mechanism is adopted to facilitate subsequent querying and analysis.

[0133] Knowledge inference is the process of deriving new knowledge based on known facts. In knowledge graphs, commonly used inference methods include rule-based inference, statistical inference, and representation learning-based inference. Rule-based inference uses predefined logical rules for deductive reasoning; statistical inference utilizes probabilistic models and statistical laws to discover potential relationships; and representation learning-based inference predicts missing links in the graph by learning low-dimensional vector representations of entities and relationships. Through knowledge inference, missing information in knowledge graphs can be supplemented, improving the completeness of knowledge.

[0134] Redundancy elimination is the process of identifying and removing duplicate or redundant information in a knowledge base. Redundancy can stem from various reasons, such as duplicate entities, duplicate relations, or logical redundancy. By employing methods such as entity alignment, relation normalization, and logical simplification, redundant information in the knowledge base can be reduced, improving the simplicity and consistency of the knowledge.

[0135] Distributed storage is a technology that distributes data across multiple physical nodes, improving storage capacity, read / write performance, and reliability. It is essential for processing large-scale knowledge graphs. Commonly used distributed storage solutions include Hadoop Distributed File System (HDFS), distributed key-value stores (such as Redis and Cassandra), and distributed graph databases (such as Neo4j and JanusGraph).

[0136] Indexing mechanisms are a key technology for improving data query efficiency. In knowledge graphs, commonly used index types include entity indexes, relation indexes, attribute indexes, and path indexes. Entity indexes accelerate queries based on entity IDs or attributes; relation indexes accelerate queries based on relation types; attribute indexes accelerate queries based on attribute values; and path indexes accelerate queries based on path patterns. A well-designed index can significantly improve the query performance of knowledge graphs.

[0137] JSONL (JSON Lines) format is a text format for storing structured data, where each line contains a separate JSON object. This format is particularly suitable for storing entities and relationships in knowledge graphs because it combines the flexibility and readability of JSON with the ease of line-by-line processing and parallel computation. In this paper, the knowledge graph is converted to JSONL format, where each entity or relationship is represented as a JSON object containing its ID, type, attributes, and connection information.

[0138] This sub-step outputs a high-quality structured knowledge base, where knowledge is stored in JSONL format and equipped with an efficient indexing mechanism. This knowledge base contains complete, accurate, and consistent structured knowledge of project success cases, providing a solid data foundation for subsequent feature extraction and multidimensional analysis.

[0139] In this embodiment, step S102 includes the following sub-steps:

[0140] Step S301: For the JSONL format project data in the structured knowledge base, apply a JSON parser to parse each project data into an independent tree structure, preserve the hierarchical relationship of key-value pairs, assign a unique identifier to each tree node, ensure that the original project can be traced in subsequent processing, and output a set of project tree structures.

[0141] This sub-step first requires processing the JSONL (JSON Lines) formatted data in the knowledge base. JSONL is a format that facilitates the storage of structured data; each line contains an independent JSON object, making it suitable for representing the complex attributes of project success cases.

[0142] JSON (JavaScript Object Notation) is a lightweight data-interchange format that organizes data in key-value pairs, supports nested structures, and naturally forms a tree-like hierarchy. In JSON, objects (represented by curly braces {}) and arrays (represented by square brackets []) constitute the hierarchical structure: object properties can be simple values ​​(such as strings and numbers) or nested objects / arrays, and arrays can contain multiple elements, which can be simple values ​​or complex objects.

[0143] A JSON parser is a tool that converts JSON text into an in-memory object structure. This paper employs a streaming JSON parser to read the JSON file line by line. This method is more memory-efficient than loading all the data at once, making it particularly suitable for handling large-scale datasets. Common JSON parsers include DOM parsers (which build the entire JSON into an in-memory tree structure) and SAX parsers (which stream and do not build a complete tree structure). The appropriate parsing strategy is chosen based on the data size and memory constraints.

[0144] A tree structure is a hierarchical data structure composed of nodes and edges, where each node (except the root node) has exactly one parent node and can have zero or more child nodes. In this example, each JSON object is parsed into a tree structure, where the root node represents the entire object and the child nodes represent individual attributes or array elements. This tree structure preserves the hierarchical relationship of the original JSON data, facilitating subsequent feature extraction and analysis.

[0145] To ensure that each node can be traced back to its original project during subsequent processing, a unique identifier is assigned to each node in the tree. Identifiers are typically designed using path representation; for example, the node corresponding to " / applying company / name" represents the value under that path. Simultaneously, a mapping table is maintained to associate each project's ID with its tree structure, ensuring that the origin of different nodes can still be distinguished after merging multiple project tree structures.

[0146] This sub-step outputs a collection of project tree structures, containing tree representations of all project data and their unique identifier mappings. These tree structures preserve all hierarchical relationships and attribute values ​​from the original JSON data, laying the foundation for subsequent construction of merged tree representations.

[0147] Step S302: Based on the project tree structure set, apply the tree merging algorithm to merge the tree structures of multiple projects into a unified merged tree representation. During the merging process, retain the independent identity information of each project, and distinguish the same paths of different projects by introducing identifier nodes and path prefixes, and output the merged tree representation.

[0148] In this sub-step, a unified merged tree representation is constructed based on the multiple independent project tree structures output in the previous step. This is one of the core innovations of this invention, aiming to merge multiple trees with similar structures into a single tree structure while preserving the independent identity information of each original tree.

[0149] Merged tree representation is a data structure that allows multiple similar tree structures to share common parts while retaining their unique information. This merging is not a simple tree concatenation, but a process of structure sharing and optimization that reduces storage redundancy and improves query efficiency.

[0150] The tree merging algorithm is the core technique for constructing merge tree representations. It first builds an empty merge tree, then processes each item tree structure sequentially, merging them into the merge tree. The merging process follows these rules:

[0151] Merging nodes with the same path: If nodes with the same path already exist in the merge tree, a new node is not created; instead, project identification information is added to the existing node.

[0152] Adding nodes for different paths: If it is a new path, create the corresponding node in the merged tree.

[0153] Project identity preservation: By storing a list of project identifiers on the node, it records which projects contain the node.

[0154] Path prefix introduction: To distinguish the same path in different projects, add a project identifier prefix to the node identifier.

[0155] For example, for two project data:

[0156] Project 1: {"Project ID": "P2023001", "Applicant Company": {"Name": "Technology Innovation Company"}};

[0157] Project 2: {"Project ID": "P2023002", "Applicant Company": {"Name": "Intelligent Company"}}.

[0158] The merged tree structure might look like this:

[0159] - Root node;

[0160] |- Project ID [P2023001, P2023002];

[0161] - Applicant companies;

[0162] |- Name [P2023001: Technological Innovation Company, P2023002: Intelligent Company].

[0163] Note that the "Name" node stores two distinct values, distinguished by the project ID as a prefix. This approach allows all data for multiple projects to be represented within a unified tree structure while maintaining the independence of the data for each project.

[0164] To improve efficiency, the following optimizations will also be performed during the merging process:

[0165] Shared structure: Identical subtree structures are stored only once, reducing memory usage.

[0166] Path compression: Long chain paths (a chain of nodes with only one child node) can be compressed.

[0167] Frequency statistics: Record the frequency of each path appearing in all projects to provide heuristic information for subsequent substructure searches.

[0168] This sub-step outputs a complete merged tree representation that efficiently stores the structured data of all items while preserving the differences between items. This representation provides an ideal data organization for subsequent efficient index building and substructure search.

[0169] Step S303: Apply the extended Burrows-Wheeler transformation algorithm to the merged tree representation to convert the tree structure into an index data structure containing three parts: a node label array, a tree topology array, and an item identifier array. This supports fast tree navigation and sub-path querying, and outputs the extended Burrows-Wheeler transformation index structure.

[0170] In this sub-step, based on the merged tree representation constructed in the previous step, an efficient index data structure is created using the extended Burrows-Wheeler transform (xBWT) algorithm. This step draws on the technique of "concise data structures based on extended Burrows-Wheeler transform".

[0171] The Burrows-Wheeler Transform (BWT) was originally an algorithm for text compression and indexing, while the Extended Burrows-Wheeler Transform (xBWT) is its extended application to tree structures. The core idea of ​​the xBWT algorithm is to transform the tree structure into several compact arrays, which together provide efficient navigation and querying capabilities for the tree.

[0172] Before applying xBWT, a depth-first traversal is first performed on the merge tree, and each node is assigned a unique preorder number. Depth-first traversal is a graph traversal algorithm that starts from the root node, explores a branch as deeply as possible, then backtracks to the previous node and explores the next branch. Preorder numbers are assigned according to the order in which nodes are visited; this numbering method helps determine the ancestor-descendant relationships between nodes.

[0173] Based on depth-first traversal and preordered numbering, construct the three main arrays of xBWT:

[0174] The node label array (Sα) stores the labels (like keys in JSON) of all nodes in the tree, arranged in a specific order. This array provides semantic information about the nodes and supports label-based queries.

[0175] Tree topology array (Sπ): Encodes the structural information of the tree, representing the position and relationship of each node in the tree. This array is usually represented by a bit vector or other compact representation, recording the parent-child relationships and hierarchical structure of nodes.

[0176] The project identifier array (Sid) records the project identifier information associated with each node, used to track the original project to which the node belongs. This array enables the differentiation of data from different projects during querying and navigation.

[0177] For example, for a simplified merge tree:

[0178] - root;

[0179] |- Project ID [P1, P2];

[0180] - Applicant companies;

[0181] |- Name [P1: Company A, P2: Company B];

[0182] |- Size [P1: Large, P2: Medium].

[0183] The following xBWT representation may be generated:

[0184] Sα = ["root", "Project ID", "Applicant Company", "Name", "Scale"];

[0185] Sπ = [0, 1, 1, 2, 2] / / Represents the index of the parent node of each node in Sα;

[0186] Sid = [{}, {P1, P2}, {}, {P1:Company A, P2:Company B}, {P1:Large, P2:Medium}].

[0187] In addition to the basic xBWT representation, auxiliary index structures will also be constructed, such as:

[0188] Path index: Quickly locate nodes on a given path.

[0189] Value index: Quickly find nodes based on their values.

[0190] Ancestor-descendant relationship index: efficiently determine the hierarchical relationship between nodes.

[0191] These index structures collectively support the following core operations:

[0192] Subtree navigation: Quickly access the child nodes from any node.

[0193] Ancestor lookup: Determines the ancestor relationship between any two nodes.

[0194] Path lookup: Quickly locate nodes based on a path.

[0195] Substructure matching: Efficiently identifies subtrees that conform to a specific pattern.

[0196] Because the data structure generated by xBWT transformation is very compact and supports efficient operations, it is particularly suitable for processing large-scale project data. Compared with traditional methods, this index structure significantly reduces memory usage while improving query speed, especially in scenarios that require frequent substructure searches.

[0197] This sub-step outputs the complete xBWT index structure, providing an efficient data access mechanism for the next step of feature extraction.

[0198] Step S304: Based on the extended Burrows-Wheeler transform index structure, combined with path decomposition, ancestor calculation and adaptive tree identifier collection technology, achieve efficient substructure search of project data, extract key features including technical field, innovation points, team size and funding needs from project data through predefined feature patterns, and output project feature set.

[0199] In this sub-step, based on the xBWT index structure built in the previous step, efficient substructure search and feature extraction are achieved. This step combines "path decomposition, ancestor calculation, and adaptive tree identifier collection" techniques, avoiding the detailed tree traversal process in traditional methods.

[0200] Substructure search refers to the process of finding substructures within a tree or graph structure that conform to a specific pattern. In project data analysis, substructure search is used to locate data segments containing specific characteristic information, such as technical field, innovation points, team size, funding needs, etc. Traditional methods require traversing the complete tree structure of each project, which is computationally expensive, while methods based on xBWT indexes can significantly improve search efficiency.

[0201] Path decomposition is a technique that breaks down complex query paths into a series of simpler path segments. For each feature pattern, it is decomposed into a series of path queries. For example, when searching for the feature "Applicant Company / Industry", it can be decomposed into searching the "Applicant Company" node, and then searching the "Industry" node among its child nodes. This decomposition method simplifies the query process and improves efficiency.

[0202] Ancestor computation is the process of determining the ancestor-descendant relationship between two nodes. In the xBWT index, ancestor relationships can be quickly determined by comparing the preorder digits and range information of nodes, without traversing the tree structure. For complex feature patterns (such as patterns involving multiple node relationships), the ancestor-descendant relationship index in xBWT efficiently determines the hierarchical relationship between nodes, avoiding repeated traversals in traditional methods.

[0203] Adaptive Tree Identifier Collection (ATC) is an optimization technique that automatically selects the most efficient identifier collection strategy based on the characteristics of the current query. It utilizes the item identifier array (Sid) in the xBWT index structure to quickly identify all items containing a specific path. This process does not require traversing the tree structure of each item; instead, it directly locates the relevant node through the index, significantly improving the efficiency of feature extraction.

[0204] Feature extraction is the process of extracting valuable features from raw data for prediction or analysis. In this paper, feature extraction is based on predefined feature patterns to extract key information from project data. First, a set of feature patterns is defined to describe the project features of interest, such as technical field, innovation points, team size, funding needs, etc. Then, for each feature pattern, the aforementioned techniques are used to efficiently search for matching substructures and extract feature values ​​from them.

[0205] Different extraction rules are used for different types of features:

[0206] Text features, such as project descriptions and technical fields, may require further text processing, such as word segmentation and keyword extraction.

[0207] Numerical characteristics, such as funding needs and team size, can be directly extracted and may be standardized.

[0208] Classification features, such as industry category and company size, may require encoding or mapping.

[0209] Finally, all extracted features are organized into structured feature vectors, with one feature vector corresponding to each item. For missing features, imputation techniques may be applied or default values ​​may be set to ensure the integrity of the feature vectors.

[0210] like Figure 2 As shown, step S103 includes the following sub-steps:

[0211] Step S401: Based on the project feature set, extract time attributes including project application time, approval time, and execution cycle; apply time series analysis algorithms to calculate the time distribution, seasonal fluctuations, and long-term trends of the number of projects, as well as the time patterns of regulatory policy changes, and output a time dimension feature vector.

[0212] In this sub-step, based on the project feature set output from the previous steps, we focus on extracting time-related features and analyzing the distribution and changing patterns of projects over time. Time-dimensional analysis is crucial for understanding historical trends in project success rates and predicting future changes.

[0213] First, extract time-related basic attributes from the project feature set. Time attributes are data describing the time point of an event or its duration. In project analysis, these mainly include: project application date (the date the project application is formally submitted), approval date (the date the project is approved or rejected), execution period (the planned implementation time of the project), and completion date (the date the project is completed). These time attributes together constitute the project's timeline, reflecting the entire process from application to completion.

[0214] Time series analysis is a statistical method used to analyze data points arranged in chronological order to identify patterns, trends, and periodic changes. In this paper, time series analysis is mainly applied to the following aspects:

[0215] Time distribution analysis of project quantity: Divide the timeline into appropriate time units (such as months, quarters, or years) and count the number of projects submitted and the number of successful projects within each time unit. This method allows for a direct observation of the time distribution of project submissions and success rates, such as a significant increase in the number of submissions or a marked improvement in the success rate during certain periods.

[0216] Seasonal fluctuation analysis: Project applications and approvals often exhibit seasonal fluctuations, with peak application periods occurring at the beginning or end of the fiscal year. Applying seasonal decomposition algorithms (such as X-12-ARIMA or STL decomposition) breaks down time series data into trend, seasonal, and residual components to identify seasonal patterns. Seasonal decomposition is the process of separating time series data into multiple components, which helps in understanding periodic changes within the data.

[0217] Long-term trend extraction: This involves using techniques such as moving averages or exponential smoothing to extract long-term trends in project application volume and success rate. Moving averages calculate the average of data points within a sliding window, smoothing out short-term fluctuations and highlighting long-term trends. Exponential smoothing, on the other hand, applies a weighted average of historical data points, giving higher weights to more recent data points, making it suitable for capturing gradual trends. These methods help identify the overall direction of project development, such as continued growth, stability, or decline.

[0218] Temporal Characteristics of Regulatory Policy Changes: Analyzing the relationship between document release dates and changes in project success rates helps identify key policy changes and their impact. This analysis can be achieved using change point detection algorithms (such as PELT or CUSUM). Change point detection is a technique for identifying points in time series data where significant changes occur, helping to uncover the impact of policy changes on project success rates.

[0219] In addition to the above analysis, a series of time-based derived features will be generated, such as whether the application was submitted during a policy peak period (Boolean feature), the time since the last relevant policy update (continuous feature), the season of application (categorical feature: spring, summer, autumn, winter), and whether the application was submitted at the end of the fiscal year (Boolean feature). These derived features further enrich the time dimension representation of the project.

[0220] All these time-dimensional features are organized into a structured time feature vector, with one vector corresponding to each project. This time feature vector comprehensively describes the characteristics of a project in the time dimension, providing important input for subsequent multi-dimensional feature fusion and helping to build a more accurate project success rate prediction model.

[0221] Step S402: Based on the project feature set, extract spatial attributes such as the geographical distribution, industry distribution, and scale distribution of the enterprises to which the project belongs. Apply spatial clustering and distribution analysis algorithms to calculate the differences in project success rates in different regions and industries, as well as the correlation between enterprise scale and project success, and output spatial dimension feature vectors.

[0222] In this sub-step, based on the project feature set, spatially relevant features are extracted to analyze the distribution characteristics and success patterns of projects across different geographical regions, industries, and other "spatial" dimensions. Spatial dimension analysis can reveal the relationship between project success and its environment, providing important environmental contextual information for the predictive model.

[0223] First, extract regional distribution characteristics. Regional distribution refers to the geographical distribution of projects or enterprises, reflecting the impact of regional factors on project success. Extract region-related attributes from the project feature set, such as the registered location, actual operating location, and project implementation location of the enterprise to which the project belongs. Standardize this regional information into unified administrative division codes for easier subsequent analysis. Based on the standardized regional information, calculate features such as regional success rate (historical project success rate for each region), regional policy inclination (the ratio of the average funding intensity of approved projects in the region to the national average), and regional development index (a comprehensive indicator combining GDP, innovation index, talent density, etc.).

[0224] Secondly, industry distribution characteristics are extracted. Industry distribution refers to the distribution of projects or enterprises across different industrial sectors, reflecting the impact of industry factors on project success. Industry-related attributes, such as industry code and sub-industry, are extracted from the project feature set and standardized into a unified industry classification standard. Based on the standardized industry information, features such as industry success rate (historical project success rate for each industry), industry policy support (frequency of mentions of each industry in policy documents and policy orientation), industry development cycle (determining the development stage of each industry), and industry popularity index (constructed by analyzing media reports, policy documents, and investment trends) are calculated.

[0225] Next, extract the characteristics of enterprise size distribution. Enterprise size is a quantitative representation of enterprise size, usually determined based on indicators such as number of employees, registered capital, and annual turnover. Extract attributes related to enterprise size and classify enterprises into large, medium, small, and micro enterprises according to the SME classification standards. Based on enterprise size information, calculate features such as size success rate (historical project success rate of enterprises of different sizes), size matching degree (the degree of matching between project type and enterprise size), and size conversion efficiency (historical efficiency of enterprises of different sizes in converting projects into actual results).

[0226] Spatial clustering is a technique for grouping objects with similar spatial characteristics, helping to identify groups of projects sharing common features. Combining geographical, industry, and scale dimensions, multidimensional clustering algorithms (such as K-means or DBSCAN) are applied to divide projects into different spatial clusters. K-means is a commonly used clustering algorithm that iteratively optimizes the division of data points into K clusters, ensuring each data point belongs to the nearest cluster center. DBSCAN, on the other hand, is a density-based clustering algorithm capable of discovering clusters of arbitrary shapes and handling noisy points. Through cluster analysis, high-success-rate clusters and low-success-rate clusters can be identified, their characteristic differences analyzed, and the distance between new projects and the cluster centers calculated as predictive features.

[0227] Distribution analysis is a technique that uses statistical methods to reveal the distribution of features across different categories. It analyzes the correlations between spatial features, such as whether certain industries have significantly higher success rates in specific regions than in other regions, or whether firms of a certain size are more likely to succeed in a particular industry. This analysis can uncover complex interaction effects and improve predictive accuracy.

[0228] Based on the above analysis, a complete set of spatial dimension feature vectors is output, containing the feature representations of all projects in the spatial dimension. These spatial dimension features, together with the time dimension features and enterprise strength features, will constitute the multidimensional feature representation of the projects, providing comprehensive feature support for success rate prediction.

[0229] Step S403: Based on the project feature set, extract strength indicators including the company's asset size, R&D investment, team composition, and historical performance. Apply feature engineering techniques to calculate the comprehensive score and hierarchical classification of the company's strength, analyze the correspondence between the company's strength and project success, and output the company's strength feature vector.

[0230] In this sub-step, based on the project feature set, we conduct an in-depth analysis of characteristics related to the company's strength. These characteristics directly reflect the applicant's comprehensive capabilities and resource endowments, and have a significant impact on the project's success rate. The analysis of the company's strength dimension can assess the project's feasibility and success potential from the perspective of the implementing entity.

[0231] First, extract asset size characteristics. Asset size is a measure of the total resources a company owns, typically expressed as asset value, capital amount, or market value. Extract asset-related information from the project feature set and external data sources (such as corporate credit information), including registered capital (the initial capital amount at the time of the company's establishment), total assets (the total value of the company's assets), net assets (the value of assets after deducting liabilities), annual operating revenue (the company's annual operating income), and profit level (a profitability indicator of the company). Standardize these numerical characteristics and calculate derived features, such as asset growth rate (the average annual growth rate of assets over the past 3 years), debt-to-equity ratio (the ratio of total liabilities to total assets), profit margin (the ratio of profit to revenue), and asset size quantile (the company's percentile ranking in the industry).

[0232] Secondly, extract the characteristics of R&D investment. R&D investment is an important indicator for measuring a company's innovation capability, reflecting the company's emphasis on technological innovation and product development. Extract indicators such as R&D investment amount, R&D investment ratio (the proportion of R&D investment to operating revenue), number of R&D personnel, R&D personnel percentage (the proportion of R&D personnel to all employees), number of patents (the number of invention patents, utility model patents, and design patents owned by the company), and number of software copyrights. Based on these basic indicators, calculate derived characteristics, such as the R&D efficiency index (the ratio of the number of patents to R&D investment), innovation activity (the growth rate of the number of newly filed patents in the past three years), technology field coverage (the number of technology fields covered by the company's patents), and core technology strength (the proportion of core technology patents in all patents).

[0233] Next, extract team composition characteristics. Team composition describes the personnel composition and capability distribution of an enterprise or project team, directly affecting the project's execution quality. Extract indicators such as team size (total number of project team members), education level distribution (educational background composition of team members), professional background distribution (professional fields of team members), experience level (average years of work experience of team members), and number of core members (number of core members with key skills or experience). Based on these fundamental indicators, calculate derived characteristics such as team fit (the degree of matching between team professional backgrounds and project fields), team stability (average length of service of core members within the enterprise), team diversity index (the degree of diversity in the professional backgrounds of team members), and high-level talent density (the proportion of personnel with senior professional titles or higher education).

[0234] Finally, extract historical performance characteristics. Historical performance represents a company's past project experience and achievements, and is an important indicator of its capabilities. Extract indicators such as the number of historical projects, historical project success rate, historical project scale, awards received, and benchmark projects. Based on these fundamental indicators, calculate derived characteristics, such as the experience accumulation index (weighted by the number of historical projects, with newer projects having higher weights), the success accumulation index (weighted by the number of historically successful projects), project complexity adaptability (an assessment of the company's ability to handle projects of varying complexity), and domain focus (the degree to which the company's projects are concentrated in a specific domain).

[0235] Feature engineering is the process of extracting, transforming, and constructing features from raw data to improve the predictive performance of a model. In this paper, feature engineering techniques are used to construct a comprehensive scoring model of enterprise strength. This model can employ various methods, such as weighted summation (assigning weights to different features and calculating a weighted total score), principal component analysis (extracting principal components to construct a low-dimensional representation of strength), and hierarchical scoring (dividing enterprise strength into levels such as financial strength, innovation capability, and team strength, scoring them separately, and then synthesizing the results). It can also construct hierarchical classifications of enterprise strength, such as categorizing enterprises into "leading enterprises," "growth-stage enterprises," and "startups," with each category potentially having different success rate features.

[0236] Based on the above analysis, a complete set of enterprise strength feature vectors is output, containing the strength feature representations of all project applicants. These enterprise strength features, together with time-dimension and spatial-dimension features, constitute a multi-dimensional feature representation of the project, providing comprehensive feature support for success rate prediction. In-depth analysis of enterprise strength enables a more accurate assessment of the project's execution capability and success potential.

[0237] Step S404: For the time dimension feature vector, spatial dimension feature vector, and enterprise strength feature vector, apply a feature fusion algorithm to weight and fuse the features of the time dimension, spatial dimension, and enterprise strength dimension, taking into account the correlation and importance between features of different dimensions, and generate the multidimensional feature representation.

[0238] In this sub-step, the feature vectors extracted from the previous three sub-steps (S401 time dimension features, S402 spatial dimension features, and S403 enterprise strength features) are intelligently fused to construct a unified multi-dimensional feature representation. The goal of feature fusion is to organically combine information from different dimensions while considering the interactions between dimensions to form a more predictive comprehensive feature.

[0239] First, correlation analysis is performed on features across all dimensions to understand the dependencies and redundancy among them. Correlation analysis is the process of assessing the strength of the statistical relationship between two or more variables. Pearson correlation coefficients or Spearman rank correlation coefficients are calculated to identify highly correlated feature pairs. The Pearson correlation coefficient measures linear correlation and is suitable for continuous variables; the Spearman rank correlation coefficient measures monotonic relationships and is suitable for any type of variable. A feature correlation heatmap is also constructed to visually display the strength of the correlation between features, and variance inflation factor (VIF) analysis is applied to detect multicollinearity issues among features. Multicollinearity refers to a high degree of correlation between predictor variables, which can lead to model instability and difficulty in interpretation.

[0240] Based on the correlation analysis results, new features reflecting the interactions between dimensions are constructed. Interactive features are new features created by combining two or more original features, capable of capturing non-linear relationships between features. Various types of interactive features are constructed: second-order interactive features (combining features from two dimensions to generate new features, such as "industry popularity × regional policy inclination"), conditional features (constructing features for one dimension based on conditions of another dimension, such as "number of manufacturing projects applied for during policy peaks"), ratio features (calculating the ratio of features from different dimensions, such as "R&D investment ratio / industry average R&D investment ratio"), and difference features (calculating the difference between actual and expected values, such as "actual success rate of enterprises - average success rate of the region"). These interactive features can capture more complex relationship patterns and improve the predictive ability of the model.

[0241] Feature importance assessment and selection is the process of determining which features are most valuable for prediction. Not all features are equally important for prediction; feature importance needs to be assessed and selected. Feature selection algorithms are applied, such as tree-based importance scoring (using decision trees or random forest models to calculate feature importance), LASSO regularization (compressing the coefficients of unimportant features to zero through an L1 penalty term), or recursive feature elimination (RFE) (iteratively training the model and removing the least important features). For each feature, information gain or Gini impurity reduction is calculated to evaluate its predictive power on the target variable. The predictive performance of different feature subsets is evaluated through cross-validation, and the optimal feature combination is selected. It may be found that certain cross-dimensional combined features (such as "corporate R&D efficiency × industry development cycle") are more predictive than single-dimensional original features.

[0242] After selecting important features, they are weighted and normalized. Weighting assigns different weights based on feature importance, with more important features receiving higher weights. Normalization makes features of different scales comparable; common methods include Min-Max normalization (scaling features to the [0,1] interval) or Z-score standardization (making the feature mean 0 and standard deviation 1). For categorical features, One-Hot encoding (converting each category into a binary feature) or target encoding (replacing category values ​​with statistics of the target variable) is applied to convert them into numerical representations.

[0243] Multiple feature fusion strategies are implemented, and the optimal solution is selected based on actual results. Simple concatenation directly concatenates feature vectors from different dimensions into a longer vector; weighted fusion is based on the predictive power of each dimension, weighting the dimensional levels; hierarchical fusion first selects and combines features within each dimension, and then fuses them across dimensions; model fusion trains a separate model for each dimension, and then fuses the model prediction results.

[0244] For high-dimensional feature spaces, representation learning techniques can be applied to reduce dimensionality and capture essential features. Representation learning is a method for automatically discovering useful representations of raw data. Common techniques include Principal Component Analysis (PCA) (extracting linearly uncorrelated principal components), autoencoders (learning low-dimensional representations of features through neural networks), and t-SNE or UMAP (preserving local relationships between features while performing nonlinear dimensionality reduction). These methods are particularly suitable for handling situations with complex nonlinear relationships between dimensions.

[0245] After the above processing, a unified multidimensional feature vector is generated. This vector is organized in a structured form, including original key features (retaining the core original features of each dimension), fusion features (new features generated through inter-dimensional interactions), dimensionality reduction features (low-dimensional representations obtained through representation learning), and meta-features (features that describe the properties of the features themselves, such as the reliability and timeliness of the features).

[0246] This multidimensional feature vector not only contains the core information of each dimension, but also captures the complex interaction relationships between dimensions, providing a comprehensive and rich feature representation for subsequent project-platform allocation models and success rate prediction models.

[0247] In this embodiment, step S104 includes the following sub-steps:

[0248] Step S501: Based on the multidimensional feature vector, apply a clustering algorithm to divide the projects into different groups according to dimensions including industry category, technical field, and enterprise size, assign weight coefficients to each group, and output the project group division results.

[0249] In this sub-step, projects are divided into different groups based on the multidimensional feature vectors generated in the previous steps. Project grouping is fundamental to achieving equitable resource allocation, as different types of projects may require different evaluation criteria and support policies.

[0250] First, identify the key dimensions for project segmentation. These dimensions are characteristics or attributes used to distinguish different project groups. In this context, they primarily include: industry category (e.g., manufacturing, services, agriculture), technology field (e.g., artificial intelligence, biotechnology, new materials), enterprise size (e.g., large enterprises, SMEs, micro-enterprises), regional type (e.g., developed regions, underdeveloped regions, special policy zones), and innovation type (e.g., basic research, applied research, technology transfer). The selection of these dimensions needs to consider policy guidance, fairness in resource allocation, and actual business needs.

[0251] A multi-level partitioning method is employed, dividing projects into primary groups and subgroups. Multi-level partitioning is a hierarchical classification method, progressively refining the classification from coarse to fine. The primary partitioning is based on the most core dimension (such as industry category) to create primary groups; the secondary partitioning, within the primary groups, further divides them into subgroups based on secondary dimensions (such as company size); there may also be overlapping groups, where some projects belong to multiple groups simultaneously, such as the "SMEs + High-Tech" overlap group. This multi-level partitioning method allows for more precise differentiation of project characteristics, providing a foundation for subsequent fairness constraint design.

[0252] In addition to partitioning based on predefined rules, clustering algorithms are applied for data-driven group discovery. Clustering algorithms are unsupervised learning methods that can automatically discover natural groupings in data based on data characteristics. Several clustering algorithms are used: K-means clustering (based on the Euclidean distance of multidimensional feature vectors, clustering items into K groups), hierarchical clustering (constructing a hierarchical structure of items, allowing for truncation at different levels to obtain groups of different granularities), density clustering (such as DBSCAN, identifying naturally formed item groups based on feature space density), and spectral clustering (handling non-convex item groups, suitable for situations with complex relationships between features). Clustering results are evaluated using metrics such as silhouette coefficient and Davies-Bouldin index to ensure the rationality of the partitioning. The silhouette coefficient measures the similarity of a point to its own cluster relative to the similarity of other clusters; a higher value indicates better clustering. The Davies-Bouldin index measures the ratio of intra-cluster distance to inter-cluster distance; a lower value indicates better clustering.

[0253] For each formed cluster, a feature analysis is performed to understand its characteristics. Feature analysis includes calculating the cluster's centroid (i.e., the average value of each feature), identifying the cluster's principal features (the features that have the highest discriminative power for that cluster), analyzing the variance and dispersion within the cluster, and calculating the distance and overlap between clusters. These analyses help to understand the characteristics and boundaries of different clusters, providing a basis for subsequent fairness constraint design.

[0254] Based on policy guidance and strategic needs, weight coefficients are assigned to each group. These weight coefficients reflect the group's priority and importance in resource allocation. Priority development groups (those aligned with national strategies and key development directions) receive higher weights; support-oriented disadvantaged groups (groups with relatively scarce resources but development potential) receive supportive weights; and balanced development groups (groups requiring balanced development) receive medium weights. These weight coefficients will influence the design of fairness constraints in subsequent steps; high-weight groups will receive more consideration in resource allocation.

[0255] Furthermore, a relationship network between groups is constructed to represent the similarity and association between groups. This group relationship network is a graph structure where nodes represent groups and edges represent relationships between groups. The similarity matrix between groups is calculated, sets of interconnected groups are identified, and the hierarchical structure and inclusion relationships between groups are analyzed. This network representation helps to understand the complex relationships between groups and supports fairness constraints across groups.

[0256] The above process outputs complete project grouping results, including the group identifier of each project, characteristic descriptions of each group, group weight coefficients, and the relationship network between groups. These results will serve as the basic input for the group fairness constraint modeling in step S503.

[0257] Step S502: Based on historical platform statistics and fairness allocation standards, determine the service capacity limit of each platform and the acceptance ratio requirement of each group. Apply convex function modeling technology to construct a convex cost function of the overall platform constraint, so that when the total number of projects allocated to the platform exceeds the soft constraint, a non-linearly increasing cost penalty will be generated, and the platform constraint model will be output.

[0258] In this sub-step, a service capacity constraint model is established for each review or funding platform based on historical data and policy requirements. Platform constraints refer to the limitations on the number and types of projects a platform can handle, directly affecting the boundary conditions of resource allocation.

[0259] First, the basic attributes of each platform are defined, including platform identifier (unique identifier and name), platform type (such as industry association, investment institution, etc.), platform function (such as policy making, fund allocation, technology evaluation, etc.), historical capacity (number and type of projects processed in the past), and resource constraints (hard resource constraints such as available funds and number of review experts). These basic attributes provide a reference for constraint modeling.

[0260] Based on historical data and current resource conditions, determine the soft capacity limit for each platform. The capacity limit is the maximum number of projects a platform can handle, categorized into soft and hard constraints. Analyze the maximum number of projects handled by the platform over historical periods, consider changes in current resource conditions (increases or decreases), incorporate policy guidance factors (some platforms may be required to expand or shrink), and ultimately set a baseline capacity (the platform's standard processing capacity) and a flexibility range (the allowable range of fluctuations). For example, platform A might have the following capacity settings: baseline capacity 100 projects / cycle, minimum capacity 80 projects / cycle (hard lower limit), maximum capacity 130 projects / cycle (hard upper limit), and flexibility range 100-130 projects (soft constraint range).

[0261] A convex function is a special mathematical function whose second derivative is always positive, and whose graph exhibits an upward convex shape. A key property of convex functions is that the function value between any two points is no higher than a linear combination of those two values, giving them favorable properties in optimization problems. In this paper, a convex function is used to model platform capacity constraints, such that when the number of items allocated to the platform exceeds the soft constraint (baseline capacity), a non-linearly increasing cost penalty occurs.

[0262] Design various types of convex cost functions to adapt to the capacity constraint characteristics of different platforms:

[0263] Quadratic cost function: When the number of projects is within the baseline capacity, the cost is zero; when the number of projects exceeds the baseline capacity but is within the hard limit, the cost increases according to a quadratic function, and the penalty increases rapidly with the increase of the excess; when the number of projects exceeds the hard limit, the cost includes a high penalty term, which strongly inhibits the allocation beyond the hard limit.

[0264] Exponential cost function: When the number of projects is within the baseline capacity, the cost is zero; when the number of projects exceeds the baseline capacity, the cost grows exponentially, providing a steeper penalty growth curve than a quadratic function, which is suitable for scenarios with stricter capacity constraints.

[0265] Piecewise linear approximation: To simplify computation, a convex function can be approximated by multiple pieces of linear functions, making it easier to use within a linear programming framework. This approximation method reduces computational complexity while preserving convexity.

[0266] Each platform will have a customized cost function form and parameters to reflect its specific capacity-cost relationship. Parameter settings will consider the platform's historical performance, current resource status, and policy orientation to ensure that the constraint model is both reasonable and flexible.

[0267] In addition to overall capacity constraints, the platform's preferences and expertise for different types of projects are also considered. A platform-project type affinity matrix is ​​constructed to allocate platform-specific processing costs to different types of projects, establishing capacity sub-constraints for platforms' areas of expertise. For example, a platform may have special expertise in the field of artificial intelligence, resulting in higher processing efficiency and capacity for such projects.

[0268] The design incorporates a dynamic adjustment mechanism to adjust platform capacity constraints based on real-time conditions. This dynamic adjustment includes periodic adjustments (updating capacity parameters based on new historical data), feedback adjustments (adjusting cost function parameters based on platform feedback), and policy adjustments (adjusting capacity allocation based on new policy guidance). This dynamic mechanism enables the platform constraint model to adapt to changing environments.

[0269] It also provides a visual representation of the constraint model for easy understanding and verification. Visualization includes plotting capacity-cost curves (intuitively showing cost changes under different project numbers), simulating total cost calculations under different allocation schemes, and comparing with historical data (verifying the model's rationality).

[0270] Through the above process, a complete platform constraint model is output, including the capacity parameters, cost function definition, and platform preference description for each platform. These models will serve as an important component of the synthesis model construction in step S504, providing key constraints for solving the optimal allocation scheme. By designing a convex cost function, the concept of "soft constraints" is realized, which allows exceeding the baseline capacity but requires paying additional costs. This is more in line with the flexibility requirements of real-world application scenarios than hard constraints.

[0271] Step S503: Based on the project group division results, apply fairness constraint modeling technology to construct a specific convex cost function for each platform and each group. This function promotes group fairness by punishing unbalanced allocation, realizes a soft constraint mechanism of restrictive dominance and minority group protection, and outputs a group fairness constraint model.

[0272] In this sub-step, based on the project grouping results of step S501, a specific fairness constraint model is constructed for each platform and group. This step directly applies the core innovation that "each platform also incurs group-specific convex costs, exceeding the number of projects that can be allocated to the platform in each group. These costs promote group fairness by penalizing imbalances, thereby generating a soft change to fairness concepts introduced in previous work, such as restrictive dominance and minority group protection."

[0273] First, two core fairness concepts are clearly defined: Restrictive Dominance and Minority Protection. Restrictive Dominance is a mechanism to ensure that no group excessively consumes resources; it prevents certain dominant groups from monopolizing resources by setting a ceiling on resource allocation. Minority Protection, on the other hand, is a mechanism to ensure that vulnerable or minority groups receive sufficient resource support; it protects the basic rights of these groups by setting a floor on resource allocation. These two concepts correspond to the upper and lower limits of resource allocation constraints, respectively, and together constitute the framework for group fairness.

[0274] Determine a reasonable quota range for each group on each platform. A quota refers to the share of resources reserved for a specific group in resource allocation. This includes three levels: minimum quota (qmin), the minimum number of projects the group should receive on the platform, achieving minority group protection; ideal quota (qideal), the ideal number of projects the group should receive on the platform, reflecting the optimal balance; and maximum quota (qmax), the maximum number of projects the group can receive on the platform, achieving restrictive control. Quota determination considers multiple factors: group size (larger groups generally should receive more resources), group weight (higher priority groups may receive higher quotas), policy orientation (groups that align with policy priorities may receive preferential quotas), and historical allocation (referencing historical allocation patterns to avoid drastic changes).

[0275] For each platform-group pair, a specific convex cost function is designed to penalize allocations that deviate from the ideal quota. A convex cost function is a function that grows non-linearly as a variable deviates from its ideal value; in this case, it is used to model fairness constraints. Several types of convex cost functions are designed:

[0276] Two-sided quadratic cost function: Costs include quadratic penalties for deviations from the ideal quota, penalties for allocations below the minimum quota, and penalties for allocations exceeding the maximum quota. The greater the deviation, the more severe the penalty, especially when allocations are below the minimum quota or above the maximum quota.

[0277] Asymmetric V-shaped cost function: The cost includes penalties for falling below the ideal quota and penalties for exceeding the ideal quota. Typically, the penalty for falling below the ideal quota is set to be greater than the penalty for exceeding the ideal quota, with a greater emphasis on the protection of minority groups.

[0278] Piecewise function: When the allocation is between the minimum and maximum quota, the cost is zero; when the allocation is below the minimum quota, the cost increases according to a quadratic function; when the allocation exceeds the maximum quota, the cost also increases according to a quadratic function. This function does not incur costs within the quota range, but only penalizes allocations that exceed the range.

[0279] In addition to quota constraints for individual groups, balance constraints are also established between groups. Inter-group balance refers to maintaining a reasonable proportional relationship between different groups in resource allocation. Several types of inter-group constraints are designed: relative proportion constraints (ensuring the allocation ratio between groups is within a reasonable range, such as the ratio of projects from high-tech enterprises to traditional enterprises should be between 1:2 and 2:1), diversity constraints (ensuring that the projects received by the platform come from a sufficiently diverse range of groups, such as any platform must receive projects from at least 3 different groups), and complementary group constraints (some groups may need to be considered collaboratively, such as maintaining a certain proportion of projects from upstream and downstream industry chain related groups).

[0280] Multiple indicators are designed to assess the fairness of allocation schemes. Fairness measures quantify the degree of fairness in resource allocation. Commonly used measures include: the Jain fairness index (measures the evenness of resource allocation; the index is 1 for perfect fairness), the minimum-maximum ratio (the ratio of the least resourced group to the most resourced group; the closer to 1, the fairer the allocation), the Gini coefficient (measures the degree of inequality in resource allocation; 0 represents perfect equality, and 1 represents perfect inequality), and the Atkinson index (a measure of inequality that considers social welfare; adjustable parameters reflect the degree of aversion to inequality). These indicators help assess and compare the fairness performance of different allocation schemes.

[0281] Unlike traditional hard quotas, this approach implements a soft constraint mechanism through a convex cost function. A soft constraint is a mechanism that allows for violation of constraints but imposes penalties, making it more flexible than a hard constraint. It allows temporary violations of quota limits, but incurs cost penalties; the cost increases non-linearly with the degree of deviation, strongly suppressing severe deviations; and it balances utility and fairness in the overall optimization objective. This soft constraint mechanism is more flexible than hard constraints, allowing for appropriate concessions in special circumstances while ensuring overall fairness through the cost mechanism.

[0282] Through the above process, a complete group fairness constraint model is output, including quota parameters for each platform-group pair, the definition of the convex cost function, and inter-group balance constraints. These models will serve as an important component of the comprehensive model construction in step S504, ensuring that resource allocation satisfies both the overall objectives and the reasonable rights and interests of each group.

[0283] Step S504: Based on the matching degree assessment of projects and platforms, define a utility value for each project-platform pair, construct an overall utility function, and combine the platform constraint model and the group fairness constraint model to construct a comprehensive optimization objective function.

[0284] In this sub-step, based on the output of the previous steps, the utility function of the project-platform is defined, and platform constraints and group fairness constraints are integrated to construct a complete project-platform allocation optimization model. This step embodies the core idea that "the cost of allocation is the sum of the values ​​of all cost functions across all groups and platforms. The goal is to find an allocation that minimizes costs while achieving at least a specified threshold of total utility for the user."

[0285] First, a utility value is defined for each project-platform pair, representing the value or benefit generated by allocating a project to a specific platform. Utility is a metric for measuring satisfaction or value in resource allocation. In this paper, the utility value is calculated using the following factors:

[0286] Match: The degree to which a project matches the platform, reflecting how well the platform suits the project type. Match can be measured by calculating the cosine similarity between the project's feature vector and the platform's preference vector; a higher similarity indicates a better match.

[0287] Project Value: The intrinsic value of a project, based on its potential impact, innovativeness, etc. Project value can be calculated by comprehensively considering dimensions such as innovativeness, potential impact, and feasibility, and using a weighted method to calculate the overall value score.

[0288] Platform Priority: The importance of a platform within the overall framework. Platform priority reflects policy direction and strategic priorities; higher-priority platforms may receive more consideration in resource allocation.

[0289] These factors are combined into a comprehensive utility calculation formula, as follows: α, β, and γ are weighting coefficients that control the relative importance of each factor. The utility values ​​of all item-platform pairs are calculated to form a utility matrix U, where U[i,j] represents the utility of item i allocated to platform j.

[0290] Based on the utility matrix, a total utility constraint is defined to ensure that the allocation scheme achieves the expected utility level. The total utility constraint takes the following form: Where x[i,j] is the decision variable, taking values ​​of 0 or 1, indicating whether to assign project i to platform j; U min This is the minimum utility threshold specified by the user. This constraint ensures that the total utility of the allocation scheme is not lower than the specified threshold, reflecting the design philosophy of ensuring utility while controlling costs.

[0291] Define a complete set of decision variables and constraints. The decision variable x[i,j]∈{0,1} represents whether project i is assigned to platform j. Basic constraints include: each project can only be assigned to one platform (…). Platform capacity constraints ,in It is the hard upper limit of platform j.

[0292] Construct the platform cost function and the group cost function. Platform cost function ,in It is the convex cost function of platform j defined in step S502; group cost function ,in It is the convex cost function of platform j with respect to group g as defined in step S503. The inter-group balance constraints defined in step S503 are also considered.

[0293] Construct a comprehensive optimization objective function: , , In addition to other constraints, this objective function embodies the core idea that "the cost of allocation is the sum of the values ​​of all cost functions across all groups and platforms." By minimizing the total cost while satisfying utility constraints, it achieves a balance between efficiency and fairness.

[0294] To facilitate subsequent solutions, the model undergoes necessary transformations and normalization. Linearization transforms the nonlinear convex cost function into a linear constraint through piecewise linearization or the introduction of auxiliary variables; slack variable introduction converts hard constraints into soft constraints with penalties; and scale normalization normalizes costs and utilities of different dimensions, allowing them to be compared on the same scale.

[0295] It provides a model evaluation mechanism to help users understand and adjust optimization objectives. Sensitivity analysis studies the impact of parameter changes on the solution, such as analysis... The impact of changes on the optimal solution; multi-objective visualization of the Pareto front of cost and utility to help understand the trade-off between the two; parameter tuning suggestions based on model behavior, providing suggestions for adjusting cost function parameters and weights.

[0296] Through the above process, a complete project-platform allocation optimization model is output, including the definitions of decision variables, utility function, cost function, constraints, and objective function. This comprehensive model will serve as the basis for solving step S105, providing a theoretical foundation for finding the optimal project-platform allocation scheme. The model implements a soft constraint mechanism through a convex cost function, allowing for flexible adjustments when necessary, while a penalty mechanism ensures the achievement of the overall objective.

[0297] In this embodiment, step S105 includes the following sub-steps:

[0298] Step S601: Based on the project-platform allocation model, construct a linear programming relaxation problem, transform the convex cost function in the original problem into a set of linear constraints through piecewise linearization, while retaining the utility constraints, thus forming a relaxation problem that can be solved by the standard linear programming method, and outputting a linear programming model.

[0299] In this sub-step, the project-platform allocation optimization model output from step S504 is converted into a linear programming (LP) relaxation problem. This is a common strategy for solving integer programming problems, as it relaxes integer constraints to obtain a problem that can be efficiently solved using standard linear programming methods.

[0300] First, the integer decision variable constraints in the original problem are relaxed to continuous variable constraints. Integer constraint relaxation refers to allowing variables that originally required integer values ​​to take fractional values. In this case, the original constraint x[i,j] ∈ {0,1} (a binary integer variable) is relaxed to 0 ≤ x[i,j] ≤ 1 (a continuous variable). This relaxation allows variables to take fractional values. For example, x[i,j] = 0.7 can be interpreted as project i having a 70% "share" allocated to platform j. Although the final actual allocation requires an integer solution, the solution to the relaxation problem provides a theoretical upper limit and heuristic information for integer solutions.

[0301] Secondly, the convex cost function in the original problem is transformed into linear constraints and an objective function. Convex cost function linearization is a technique that represents a nonlinear convex function as a set of linear constraints; in this paper, piecewise linearization is mainly used. Piecewise linearization approximates the convex function f(x) as a multi-segment linear function, each segment represented by a different slope. Specifically, the domain is divided into multiple intervals [0, d1], [d1, d2], ..., [d...]. n-1 ,d n Introduce new variables y1, y2, ..., y in each interval. n This represents the value of the variable within that interval; adding constraints ensures the interval is filled sequentially: 0 ≤ y k ≤ d k - d k-1The original variables are expressed as the sum of the new variables: x = y1 + y2 + ... + y n The cost in the objective function is expressed as: f(x) ≈ s1y1 + s2y2 + ... + s n y n S K It is the slope of the k-th segment, usually taken as the average slope of the original function in that interval or the slope of the interval boundary.

[0302] The quadratic cost function for the overall platform constraints If x>C, it can be linearized as follows: Define a variable z to represent the part that exceeds the capacity: z = max(0, xC); Add constraints: z ≥ x - C, z ≥ 0; Divide the interval [0,HC] into multiple segments and apply piecewise linearization.

[0303] For a two-sided quadratic cost function with group fairness constraints

[0304] This can be linearized as follows: Define the variable u = x - q ideal Define variable v1 = max(0, q); min Add the constraint: v1 ≥ q (-x). min -x, v1≥ 0; Define variable v2= max(0, xq max Add the constraint: v2 ≥ xq max v2≥ 0; The squared terms of u are processed by piecewise linearization.

[0305] To simplify the problem representation, aggregate variables are introduced. The total platform allocation N j = Σ i x[i,j] represents the total number of projects allocated to platform j; group-platform allocation. This represents the number of items assigned to platform j within group g. These aggregate variables make the cost function more concise and reduce the number of constraints introduced during linearization.

[0306] The utility constraint itself is already in linear form, so it can be directly incorporated into the linear programming model: , where U[i,j] is the pre-computed utility value of project i assigned to platform j.

[0307] Organize all the transformed constraints and objective function into standard linear programming form: minimize Z = c Tx, subject to Ax ≤ b, x ≥ 0, where x is a vector containing the original decision variable x[i,j] and all auxiliary variables, c is the objective function coefficient vector containing the slope of each linearized segment, A is the constraint matrix, and b is the constraint right-hand constant vector.

[0308] Since linearization can significantly increase problem size, measures are taken to control the size. The adaptive segmentation strategy uses finer segments in cost-sensitive intervals and coarser segments in intervals with moderate cost changes; problem structure utilization involves identifying and removing redundant constraints, merging similar constraints to reduce the total number of constraints, and simplifying the representation by utilizing the problem's network structure; variable pre-screening is based on utility thresholds to pre-exclude impossible project-platform assignments, and for clearly mismatched project-platform pairs, x[i,j] = 0 is directly fixed.

[0309] Verify the solvability of the linear programming relaxation problem, check the consistency of constraints to ensure there are no contradictory constraints; verify that at least one feasible solution satisfies all constraints; analyze whether the objective function is bounded. If the problem is found to be unsolvable or the solution is unstable, diagnostic information and correction suggestions will be provided.

[0310] Through the above transformation, a complete linear programming relaxation problem is output, including the standard form of decision variables, objective function, and constraints. This linear programming problem retains the core structure and constraints of the original problem, and by relaxing integer constraints and linearizing the convex cost function, it is transformed into a form that can be efficiently solved using standard linear programming algorithms, providing an important foundation for constructing integer solutions in subsequent steps.

[0311] Step S602: Solve the linear programming model using the simplex method or interior point method, obtain fractional solutions during the solution process, and output the optimal solution of the linear programming problem.

[0312] In this sub-step, the linear programming relaxation problem constructed in step S601 is solved to obtain the optimal or near-optimal solution. Linear programming is a classic optimization problem with mature algorithms and tools available, such as the simplex method and the interior point method.

[0313] The Simplex Method is an iterative algorithm that finds the optimal solution by moving between vertices in the feasible region. The optimal solution to a linear programming problem (if it exists) is always reached at a vertex in the feasible region. The Simplex Method leverages this property, starting from an initial feasible vertex and moving along the boundary to an adjacent vertex with a better objective function value, until no better vertex can be found. The basic steps of the Simplex Method include: selecting an initial basic feasible solution; calculating the reduced cost of all non-basic variables; selecting the non-basic variable corresponding to the most negative cost as the entering variable; determining the ratio of the non-basic variable to the entering variable; updating the basic feasible solution; and repeating until all cost values ​​are non-negative. Although the Simplex Method may theoretically require exponential time, it performs excellently in practice, efficiently handling sparse constraint matrices and is suitable for small to medium-sized problems (within the thousands of variables and constraints).

[0314] The interior point method is an algorithm that approximates the optimal solution by starting from within the feasible region and following a central path. Unlike the simplex method, the interior point method does not move between vertices but seeks the optimal solution within the feasible region. The basic idea of ​​the interior point method is to transform the constraints into obstacle functions and then solve a series of unconstrained optimization problems, gradually approaching the optimal solution of the original problem. The interior point method theoretically has a good guarantee of polynomial time complexity and is suitable for large-scale problems (with tens of thousands of variables and constraints), performing even better for dense constraint matrices. Common variations of the interior point method include the primal-dual interior point method, the predicate center method, and the affine scaling method.

[0315] Choose an appropriate linear programming algorithm based on the problem size and characteristics. For project-platform allocation problems, which are usually of medium size but have sparse constraint matrices, the simplex method or a hybrid strategy (first use the interior point method to quickly approach the region of the optimal solution, and then use the simplex method to obtain the accurate vertex solution) may be preferred.

[0316] To improve solution efficiency, the selected solution algorithm is configured and optimized. Preprocessing optimization includes constraint matrix compression (removing redundant constraints and merging similar constraints), variable pre-screening (identifying and fixing variables with definite values), problem reduction (reducing the size of effective variables and constraints), and numerical scaling (improving the condition number of the constraint matrix and enhancing numerical stability). Solver parameter tuning includes accuracy control (setting appropriate optimality and feasibility tolerances), iteration constraints (setting maximum number of iterations and time limits), numerical stability parameters (preventing numerical overflow and accumulation of rounding errors), and heuristic parameters (controlling variable selection rules, entry variable rules, etc.).

[0317] For large-scale problems, parallel computing strategies may be employed, including multi-threaded or distributed computing, parallel solution of subproblems, and GPU-accelerated matrix operations (such as large matrix operations in the interior-point method). For complex problems, a phased solution strategy may be adopted, such as first solving basic linear programming relaxation problems to obtain preliminary solutions as the starting point for subsequent stages; gradually adding more complex constraints, with each stage using the solution of the previous stage as a hot start point; using relatively relaxed accuracy requirements in the initial stage to quickly obtain approximate solutions, and increasing the accuracy requirements in subsequent stages to accurately locate the optimal solution.

[0318] After solving the linear programming problem, analyze its characteristics. Basic solution analysis includes identifying basic and non-basic variables, analyzing the sparsity and structural characteristics of the solution, and assessing integer properties (the number of variables naturally results in integer values). Sensitivity analysis includes calculating the allowable range of variation for the objective function coefficients, analyzing the sensitivity of the constraint right-hand side constants, and identifying bottlenecks and redundancies in the constraints. Degradation detection includes detecting whether the solution exhibits degradation, analyzing the possibility of multiple optimal solutions, and assessing the stability of the solution and its sensitivity to input perturbations.

[0319] The original solutions obtained are processed and transformed as necessary, including converting them from a solver-specific format to a unified internal representation and reconstructing the multidimensional structure of decision variables (from one-dimensional vectors back to two-dimensional matrices). Graphical representations of the solutions are also generated, such as allocation network diagrams and intuitive displays of key indicators (e.g., platform capacity utilization, group allocation ratios). For fractional solutions, the variables that take fractional values ​​and their distributions are identified, and the structural patterns of fractional solutions are analyzed to provide insights for subsequent integerization.

[0320] Through the above solution process, the optimal solution to the complete linear programming relaxation problem is output, including the values ​​of all decision variables and the objective function, as well as the quality index and sensitivity analysis results of the solution. This solution will serve as the input for the network flow model transformation in step S603. Due to the relaxation of integer constraints, the solution may contain fractional values, which will be transformed into integer solutions in subsequent steps using network flow techniques.

[0321] Step S603: Based on the optimal solution of the linear programming, construct the corresponding network flow model, regard the project as the source and the platform as the sink, convert the score allocation in the optimal solution of the linear programming into the flow in the network, construct the minimum cost flow problem that satisfies the capacity constraint and flow conservation, and output the network flow model.

[0322] In this sub-step, the optimal solution of the linear programming problem obtained in step S602 is converted into a network flow model. The network flow model is a special type of linear programming problem with good structural properties, which can be solved using efficient combinatorial optimization algorithms. This step directly reflects the core idea that "our algorithm is based on techniques involving linear programming and network flow."

[0323] First, the basic structure of the network flow model is designed, including the definition of nodes and edges. In the network flow model, nodes represent locations or entities in the network, and edges represent connections or flow paths between nodes. This network flow model contains four types of nodes: source nodes (S) represent the starting point of all items; item nodes (P) represent the starting point of all items; and item nodes (P) represent the starting point of all items. i ) represents each project; platform node (F) j () represents each platform; sink(t) represents the destination of all traffic. Edge types include: source-to-project edge (s→P) i ), with a capacity of 1, indicating that each project must be fully allocated; project to platform edge (P i →F j ), with a capacity of 1 and a cost of -U[i,j] (negative utility), representing the utility of the item allocated to the platform; the edge from the platform to the sink (F j →t), where capacity is the upper limit of platform j's capacity, and cost is the platform's capacity cost function.

[0324] Network flow problems study the allocation and optimization of traffic in a network. A flow refers to the movement of matter or information from a source to a sink, and the flow on each edge must satisfy capacity constraints and flow conservation. In the minimum cost flow problem, the goal is to find the flow allocation that satisfies the flow requirements while minimizing the total cost. This paper constructs a minimum cost flow problem, where the flow requirement is that the total flow from the source s to the sink t equals the total number of items.

[0325] The optimal solution to the linear programming relaxation problem is mapped to the initial flow in the network. For each item i and platform j, if x[i,j]>0, then on edge P i →F j The flow rate is set to x[i,j]. Since the linear programming solution may contain fractional values, the initial flow may not be an integer flow. The flow rate from source node s to each project node P... i The traffic is set to 1, indicating that each project needs to allocate a full share; each platform node F j Total inflow equals Σ i x[i,j] represents the total number of projects allocated to the platform in the linear programming solution; the total inflow to sink t equals the total number of projects, indicating that all projects are allocated.

[0326] Group fairness constraints cannot be directly represented in standard network flow models, so a network extension method is used to handle them: an intermediate node G is added to each group g. g Modify the project node P belonging to group g. i To platform node F j Change the edge to P i →G g →F j The two paths; in G g →F jAdd upper and lower capacity constraints to the edges to reflect the group quota requirements. Alternatively, the convex cost function of group fairness can be transformed into a piecewise linear cost function on the edges; for example, the portion exceeding the group's maximum quota will incur a higher unit cost. Or, the group fairness constraints can be incorporated into the objective function using the Lagrange multiplier method through Lagrange relaxation, adjusting the multiplier value so that the solution gradually satisfies the group fairness requirements during the iteration process.

[0327] Choose an appropriate solution algorithm based on the characteristics of the constructed network. The minimum-cost maximum flow algorithm is a classic method for solving minimum-cost flow problems, including augmenting path methods (finding the shortest cost augmenting path from the source to the sink and iteratively increasing the flow), cost scaling techniques (gradually approaching the optimal solution by scaling the cost of edges), and capacity scaling techniques (processing large-capacity edges first, then considering small-capacity edges). Cycle elimination algorithms start from an initial feasible flow and continuously search for and eliminate negative-cost cycles. The network simplex method is a special simplex algorithm for network flow problems, utilizing network structure to improve efficiency and suitable for handling sparse networks and network flow problems with secondary constraints.

[0328] The design mechanism ensures that an integer flow solution is ultimately obtained. The all-integer property of network flow problems states that an integer-optimal flow exists when all capacities are integers. This property is used to obtain an integer solution by ensuring that the capacity of all edges in the network is set to integers. For initial solutions containing fractional flows, approximate integerization techniques may be applied, such as random rounding, deterministic rounding, or network simplex pivot operations; alternatively, flow decomposition techniques can be used to decompose the initial flow into a set of path flows from the source to the sink, selecting the optimal combination of integer path flows to ensure that the total flow remains constant.

[0329] To improve solution efficiency, the network size is optimized. Edge filtering removes edges where x[i,j]=0 in the linear programming solution, as these edges are likely to be zero in the optimal integer solution as well. Node merging identifies nodes with the same connection pattern and merges them without changing the nature of the problem. Layered network construction divides the project into layers according to certain characteristics, reducing cross-layer connections and leveraging the hierarchical structure of the problem to reduce network complexity.

[0330] It provides a visual representation of the network model for easy understanding and analysis. The network topology diagram intuitively displays the structure of nodes and edges, using color and thickness to represent flow size and cost; the initial flow distribution diagram shows the initial flow distribution mapped from the linear programming solution to the network, marking fractional flows and integer flows; the group view groups network nodes by group and platform, highlighting the structure related to group fairness constraints.

[0331] Through the above process, a complete network flow model is output, including the definitions of nodes and edges, capacity and cost settings, initial flow allocation, and solution strategy. This network flow model retains the core structure and constraints of the original problem, while utilizing the special properties of the network flow problem to provide a theoretical guarantee and algorithmic foundation for the extraction of integer solutions in step S604. Constructing the network flow model is a key step in transforming the complex project-platform allocation problem into a standard combinatorial optimization problem that can be solved using efficient algorithms.

[0332] Step S604: Apply the minimum cost flow algorithm to the network flow model to obtain the integer flow solution, thereby determining the integer allocation scheme from the project to the platform. Through theoretical analysis, ensure that the cost difference between the approximate solution and the global optimal solution does not exceed the preset approximation ratio, while satisfying the utility constraint, and output the optimal allocation scheme.

[0333] In this sub-step, based on the network flow model constructed in step S603, integer solutions are extracted and the final project-platform allocation scheme is generated. This is the last step in the entire algorithm process, directly producing an executable decision result, realizing the core idea of ​​"ensuring, through theoretical analysis, that the cost difference between the approximate solution and the global optimal solution does not exceed the preset approximation ratio, while satisfying utility constraints."

[0334] First, the minimum cost flow algorithm is applied to the constructed network flow model to obtain the optimal or near-optimal integer flow. The minimum cost flow algorithm is the standard method for solving the minimum cost flow problem, capable of finding the optimal solution in polynomial time. Algorithms such as augmenting path methods, cost scaling techniques, cycle elimination algorithms, or network simplex methods can be selected, with the most suitable method chosen based on the network size and characteristics.

[0335] The augmenting path method starts with an initial flow and iteratively finds the minimum-cost path from the source to the sink, increasing the flow along this path until the required total flow is reached. The specific steps include: initialization (using the flow mapped from the linear programming solution as the initial solution); constructing the residual network (calculating the remaining capacity for each edge based on the original capacity and the current flow); finding the shortest path (finding the minimum-cost path from the source to the sink in the residual network); augmentation operation (increasing the flow along the found path until the minimum remaining capacity on the path is reached); updating the flow and the residual network; repeating until no path from the source to the sink can be found, or the required total flow is reached.

[0336] Cost scaling is a method to accelerate the convergence of the minimum cost flow algorithm. The basic idea is to initially set a large cost scaling factor ε, find the ε-optimal solution under the current scaling factor, and then gradually decrease ε, making the solution gradually approach the global optimum. This method avoids the slow convergence problem when the algorithm is close to the optimal solution, thus improving the algorithm's efficiency.

[0337] The network simplex method is a variant of the simplex algorithm specifically designed for network flow problems. The algorithm steps include: constructing an initial spanning tree solution (based on a linear programming solution); calculating the reduced cost of all non-tree edges; selecting the non-tree edge with the most negative reduced cost to enter the basis; determining the tree edges leaving the basis; updating the spanning tree and flow assignment; and repeating until the reduced cost of all non-tree edges is non-negative. This method leverages the characteristics of network structure, significantly reducing computational cost, and is particularly suitable for handling large-scale network flow problems.

[0338] Ensure the obtained solutions satisfy the integer property to become an executable project-platform allocation scheme. The all-integer property of network flow problems guarantees the existence of an integer-optimal solution when all capacities in the network are integers. Verify that the final flow is an integer on every edge to ensure the feasibility of the allocation scheme. For possible non-integer flows (due to numerical errors or algorithm characteristics), apply integerization, such as rounding down or up, while adjusting the flow on relevant paths to maintain flow conservation. If multiple optimal integer solutions exist, a secondary objective function may be applied to select the best scheme, such as prioritizing solutions with better group balance or solutions with more balanced platform load.

[0339] Verify whether integer solutions satisfy the key constraints of the original problem. Utility constraint verification involves calculating the total utility corresponding to the integer solutions. ), verify U total ≥ U min If the conditions are not met, it may be necessary to try local adjustments or relax other constraints before solving. Platform capacity constraint verification involves calculating the total number of projects (N) allocated to each platform j. j = Σ i x[i,j]), check if the hard capacity limit (N) is met. j ≤ C j，max ), calculate the cost penalty for exceeding soft capacity. Group fairness constraint verification is performed for each platform j and group g, calculating the number of items allocated ( ), check deviations relative to quota limits and calculate group fairness cost penalties.

[0340] Analyze the difference between the integer solution and the theoretical optimal solution to ensure it remains within an acceptable range. The approximation ratio is the ratio of the objective function value of the integer solution to the objective function value of the theoretical optimal solution, reflecting the quality of the solution. Through theoretical analysis, an upper bound for the approximation ratio is derived, proving that under specific conditions, this difference has an upper bound guarantee. Typically, it can be proven that the approximation ratio is (1+ε) or a constant factor. Calculate the objective function value (Z) of the integer solution. int ) and the objective function value (Z) of the linear programming relaxation solution lp ), calculate the actual approximation ratio (ρ = Z) int / Z lpThis verifies whether ρ is within the theoretically guaranteed range. It also evaluates the impact of small input variations on the approximation ratio, identifies key factors affecting the approximation ratio, and provides a basis for optimizing algorithm parameters.

[0341] Local search techniques can be applied to further optimize the initial integer solution. Local search is a method of finding better solutions by exploring the neighborhood of the current solution in the solution space. A swap operation attempts to exchange the allocation platforms of two items; if the total cost decreases after the swap and the constraints are still satisfied, the swap is accepted. A move operation attempts to move a single item from one platform to another; if the total cost decreases after the move and the constraints are still satisfied, the move is accepted. More advanced local search methods include tabu search (using a tabu list to avoid getting trapped in local optima) and simulated annealing (accepting adjustments that increase cost with a certain probability, gradually decreasing the probability of accepting suboptimal solutions as iterations progress).

[0342] Finally, based on the final integer solution, a complete project-platform allocation scheme is generated. The allocation scheme includes the values ​​of decision variables (for each project i, a unique allocation platform j is determined such that x[i,j]=1), allocation statistics (the total number and distribution of projects obtained by each platform, and the distribution of each group across platforms), and key metrics (platform utilization, group allocation ratio, etc.). A graphical representation of the allocation scheme is also created, displaying the project-platform allocation network and highlighting key allocation patterns and characteristics.

[0343] Through the above process, a complete project-platform final allocation scheme is output, including the allocation decision for each project, cost and utility assessment of the scheme, constraint satisfaction, and approximation ratio guarantee. This allocation scheme will serve as the final decision output, guiding the actual project allocation implementation. The allocation scheme satisfies both efficiency and fairness requirements, while also possessing theoretical quality guarantees, embodying the core value of an "efficient polynomial-time approximation algorithm."

[0344] In this embodiment, step S106 includes the following sub-steps:

[0345] Step S701: Based on the optimal allocation scheme and historical project success data, construct a model training dataset, and output the processed training dataset through data cleaning and standardization, including feature standardization, missing value handling and outlier detection.

[0346] In this sub-step, based on the previously obtained optimal allocation scheme and historical project success data, a high-quality dataset is constructed for training the prediction model. The quality of the training dataset directly affects the performance of the prediction model; therefore, the data preparation stage is crucial.

[0347] First, multi-source data is integrated to form the original training dataset. The optimal allocation scheme data includes the project-platform allocation relationship, platform evaluation information, and basic project characteristics; historical project success data includes the success status of completed projects, success rate scores, and key milestone information. These datasets are then fused using association fields such as project identifiers, company information, or technology fields to ensure that each training sample contains both input features (project features, platform features, matching features, etc.) and target variables (whether the project was ultimately successful, success rate scores, etc.). This data fusion process may face challenges such as field inconsistencies, format differences, and identifier mismatches. These issues are addressed using field mapping tables, fuzzy matching algorithms, and data validation rules to ensure the accuracy of data associations.

[0348] Data cleaning is the process of identifying and correcting errors, inconsistencies, or missing data in a dataset, and it is a fundamental step in data preprocessing. First, a format consistency check is performed to ensure that all numeric fields actually contain numbers, date fields use standard formats, and categorical field values ​​are within a predefined valid set. Then, duplicate records are detected and processed, including complete duplicates (all field values ​​are the same) and partial duplicates (key fields are the same, but other fields may differ). Complete duplicates are usually deleted directly, while partial duplicates require merging or selecting the latest record according to business rules. Data cleaning also includes error value correction, such as numerical values ​​outside the valid range (e.g., negative project durations), obviously incorrect categorical values ​​(e.g., non-existent industry codes), or malformed strings (e.g., email addresses missing the @ symbol). These errors are identified through business rule validation, pattern matching, and anomaly detection algorithms, and corrected according to predefined rules or nearest-neighbor values.

[0349] Missing value handling is a technique for resolving missing data in a dataset and is crucial for constructing a complete training set. First, the missing value pattern is analyzed to determine the type of missing value: Completely Random Missing (MCAR), Random Missing (MAR), or Non-Random Missing (MNAR). Based on the missing value pattern analysis, an appropriate handling strategy is selected. For a small number of missing important features, imputation techniques are used to fill in the missing values. Common methods include mean / median / mode imputation (replacing missing values ​​with feature statistics), nearest neighbor imputation (filling with values ​​from similar samples), regression imputation (predicting missing values ​​based on other features), or multiple imputation (generating multiple possible imputation values ​​to reflect uncertainty). For features with high missing rates, their importance is assessed; the feature may be directly deleted, or a "missing indicator" feature may be created to mark missing values. For samples with a large number of missing features, if the sample size is sufficient, these samples may be deleted; otherwise, the representativeness of the samples is assessed to decide whether to retain them and perform more complex imputation.

[0350] Outlier detection is the process of identifying observations in a dataset that significantly deviate from the normal pattern. Outliers may represent genuine anomalies or data errors. Various methods are used to identify outliers: statistical methods (such as Z-scores and IQR rules), density-based methods (such as Local Outlier Factor (LOF), clustering methods (such as DBSCAN), or model-based methods (such as Isolation Forest). For identified outliers, their legitimacy and impact are assessed: outliers confirmed as erroneous are corrected or removed; outliers confirmed as genuine but extreme may be truncated (limited to a reasonable range) or their impact mitigated through transformation; boundary outliers (close to but exceeding the normal range) may be retained but treated specially in the model. All outlier handling operations are recorded to ensure the transparency and traceability of the analysis process.

[0351] Feature standardization is the process of converting features of different dimensions and ranges to a uniform scale, and it is crucial for many machine learning algorithms. Various standardization techniques are implemented: Min-Max normalization scales features to the [0,1] or [-1,1] interval, suitable for data with unknown distributions or non-Gaussian distributions; Z-score standardization sets the feature mean to 0 and the standard deviation to 1, suitable for data with approximately normal distributions; robust standardization uses the median and interquartile range instead of the mean and standard deviation, and is insensitive to outliers; logarithmic transformation is suitable for handling highly skewed distributions, especially long-tailed distributions. Different standardization methods may be applied to different types of features: numerical features (such as project budgets and team size) typically use Min-Max or Z-score standardization; temporal features (such as project duration and application-approval intervals) may be processed through logarithmic transformation; count features (such as the number of patents and the number of historical projects) may use square root transformation or logarithmic transformation. Domain knowledge and physical meaning of the features are also considered to ensure that the standardized features remain interpretable.

[0352] For categorical features, appropriate encoding methods are applied to convert them into numerical representations. A common method is One-Hot encoding, which converts each category into a binary feature. For example, project types ("research-oriented," "development-oriented," "conversion-oriented") become three features after One-Hot encoding: "whether research-oriented," "whether development-oriented," and "whether conversion-oriented." For categorical features with high cardinality (such as industry codes), One-Hot encoding may lead to feature explosion. Alternative methods may be used, such as target encoding (replacing category values ​​with statistics of the target variable), hash encoding (mapping categories to a fixed number of buckets), or embedding encoding (learning a low-dimensional representation of the categories). For ordered categorical features (such as risk ratings: low, medium, high), ordinal encoding (assigning incrementing integers) or binary encoding (using binary bits to represent the order) may be used.

[0353] Through the data cleaning and standardization processes described above, a high-quality training dataset is output, containing complete, consistent, and normalized features and target variables. This processed training dataset provides a solid foundation for subsequent feature encoding and model building, ensuring the training quality and generalization ability of the predictive model. A data processing report is also generated, recording basic statistical information of the dataset, detailed records of the cleaning operations, and standardized parameters, providing a reference for subsequent analysis and model interpretation.

[0354] Step S702: Apply the improved folded Reed-Solomon code technique to the processed training dataset for feature encoding, mapping the high-dimensional feature space to the low-dimensional encoding space, and outputting the encoded feature data.

[0355] In this sub-step, an innovative encoding technique—improved Folded Reed-Solomon Codes (FRS)—is applied to the processed training dataset to achieve efficient feature representation. This is one of the core innovations of this invention, significantly improving the efficiency and robustness of feature encoding by introducing information theory and coding theory methods into the field of machine learning feature processing.

[0356] Reed-Solomon codes are a classic type of error-correcting code, proposed by Irving Reed and Gustave Solomon in 1960, initially used for error correction in communication and storage. They belong to the category of non-binary cyclic codes and are based on polynomial computation over finite fields. The encoding process of a traditional Reed-Solomon code can be viewed as using message symbols as polynomial coefficients and evaluating the polynomial at specific points; the decoding process reconstructs the original polynomial from these evaluated points, even if some evaluation points contain errors. This method has strong error-correcting capabilities, able to correct multiple consecutive or random errors.

[0357] Folded Reed-Solomon Codes (FRS) are an improvement on traditional Reed-Solomon Codes. They reduce the code length through a "folding" operation while maintaining strong error correction capabilities. In traditional Reed-Solomon Codes, each evaluation point generates a single coded symbol; in the folded version, multiple consecutive evaluation points are combined into a larger symbol. This folding operation increases the code rate (the proportion of information bits to total bits) while preserving the original code's error correction capabilities, making it particularly suitable for handling noisy environments. In this paper, FRS codes are innovatively applied to the field of feature coding, mapping a high-dimensional feature space to a low-dimensional coding space while retaining key information.

[0358] First, features are prepared and transformed to prepare for FRS encoding. This includes grouping and standardizing features to meet the input requirements of the encoding algorithm. Feature grouping is the process of combining related features together, which can be based on domain knowledge (e.g., grouping all time-related features together) or data-driven methods (e.g., based on correlation or clustering results). For large feature sets, dimensionality reduction techniques such as Principal Component Analysis (PCA) or autoencoders may be applied first to extract principal components, simplifying the subsequent encoding process. Furthermore, numerical features are quantized, mapping continuous values ​​to a finite set of discrete values ​​to fit the requirements of FRS encoding. Categorical features may require special processing, such as target encoding or embedding vectors, to ensure that the encoding process preserves the semantic relationships between categories.

[0359] This paper innovatively applies an improved FRS coding algorithm, mapping processed features to a low-dimensional coding space. In traditional FRS coding, the message is treated as polynomial coefficients over a finite field; however, in this improved version, feature vectors are treated as information to be encoded, mapped to codewords through a specific coding function. Several innovations are made to FRS coding: an enhanced folding mechanism, optimizing folding parameters to achieve a better balance between code rate and error recovery capability; a locality-sensitive coding function to ensure that similar feature vectors map to similar codewords, preserving the local structure of the feature space; adaptive coding parameters that dynamically adjust coding parameters based on data distribution and noise levels; and a hierarchical coding strategy that applies different levels of coding protection to features of varying importance.

[0360] The key steps in implementing FRS encoding include: First, dividing the feature vector into fixed-length segments, each corresponding to the coefficients of a polynomial; then, evaluating each polynomial on a specific set of points in a pre-selected finite field, with these evaluation results forming the encoded symbols; next, folding and combining m consecutive encoded symbols into a larger symbol, where m is a folding parameter controlling the compactness of the encoding; finally, applying a specific mapping function to convert the folded symbols into the final encoded feature representation. The entire process can be viewed as mapping the original high-dimensional feature space to a structured low-dimensional space through mathematical transformations, while preserving the essential information needed for recognition and classification.

[0361] FRS encoding offers several advantages in feature processing. First, it reduces dimensionality, compressing high-dimensional features into a more compact representation, thus reducing the complexity of subsequent calculations. Second, it enhances noise resistance, resisting noise and perturbations in the feature space and improving model robustness. Third, it preserves feature structure; the encoding process retains the key structure and relationships of the original feature space, avoiding the loss of important information. Fourth, it improves computational efficiency, especially in scenarios requiring frequent feature matching or similarity calculations, where encoded features can significantly accelerate the computation process. These advantages make FRS encoding particularly suitable for handling complex high-dimensional feature spaces in project success rate prediction.

[0362] The paper also considers potential challenges in FRS encoding and designs corresponding solutions. For the encoding loss problem (i.e., the potential loss of some information during encoding), information loss is minimized by optimizing encoding parameters, preserving the complete information of important features, and designing locally preserving encoding functions. For the computational complexity problem, an efficient encoding algorithm is implemented, utilizing parallel computing and table lookup optimization to accelerate the encoding process. For the feature heterogeneity problem (different types of features may require different encoding strategies), a hierarchical encoding framework is designed, applying different encoding parameters and strategies based on feature type and importance.

[0363] Through the above process, the output is encoded feature data, containing low-dimensional encoded representations of the original high-dimensional features. These encoded features retain the key information of the original features while possessing higher computational efficiency and noise robustness, providing ideal input for subsequent decoder design and model construction. The encoding parameters and mapping relationships are also preserved to ensure consistent encoding processing can be applied to new items during the prediction phase.

[0364] Step S703: Based on the encoded feature data, design a deterministic decoder with near-linear time complexity. The deterministic decoder can complete the decoding process in near-linear time and output a deterministic prediction model.

[0365] In this sub-step, an efficient deterministic decoder is designed based on the encoded feature data from the previous step. The decoder is the component that transforms the encoded features back to the original feature space or directly maps them to the prediction results; its efficiency directly affects the real-time performance of the prediction. This paper innovatively designs a deterministic decoder with near-linear time complexity, significantly improving prediction efficiency.

[0366] Near-linear time complexity refers to an algorithm whose time complexity is close to O(n·polylog(n)), where n is the input size and polylog(n) represents a polynomial of log(n). Compared to the quadratic or higher complexity of traditional decoders, near-linear time decoders can process large-scale data more efficiently. Deterministic algorithms are those that always produce the same output given an input. Compared to randomized algorithms, deterministic algorithms are more reliable and predictable, making them particularly suitable for scenarios with high stability requirements.

[0367] First, we analyze the mathematical structure of FRS encoding to identify optimizable decoding paths. Traditional Reed-Solomon decoding typically employs polynomial-based reconstruction methods, such as the Berlekamp-Massey algorithm or the Euclidean algorithm. These methods often require O(n) time complexity when errors occur. 2 The time complexity of this method is reduced. By analyzing the special structure of feature encoding, this paper identifies decoding paths that can be accelerated: optimizing the decoding process for specific encoding parameters, utilizing the sparsity of encoding features, applying fast algorithms to replace traditional methods, and pre-computing commonly used decoding patterns.

[0368] A decoding algorithm based on the Fast Fourier Transform (FFT) is designed to significantly improve computational efficiency. The FFT is an efficient algorithm for calculating the Discrete Fourier Transform with a time complexity of O(n log n). In this paper, the FFT is used to accelerate the polynomial evaluation and interpolation processes, which are the core steps in decoding FRS codes. First, the decoding problem is transformed into a form suitable for the FFT through preprocessing; then, the FFT is used to perform polynomial evaluation or interpolation; finally, the decoding result is obtained through post-processing. This method is particularly suitable for processing large-scale data, as its performance advantage becomes more pronounced with increasing data size.

[0369] Furthermore, sparse recovery techniques are applied to further improve decoding efficiency. Sparse recovery refers to the technique of reconstructing sparse signals from a small number of observations and is widely used in compressed sensing. Analysis of the sparsity patterns of item features revealed that most items in a specific domain occupy only a small number of non-zero values. Based on this observation, a decoding strategy based on sparse recovery was designed: first, the sparse pattern of the feature vector is estimated to identify possible non-zero locations; then, precise decoding is performed only at these locations, significantly reducing computational complexity; finally, iterative optimization improves the recovery accuracy. This method is particularly suitable for handling high-dimensional sparse features, significantly reducing computational complexity while maintaining decoding quality.

[0370] A hierarchical decoding architecture was designed, adjusting the decoding precision based on feature importance and application scenario. Hierarchical decoding refers to a method of decoding information layer by layer according to requirements, suitable for application scenarios requiring different levels of precision. Features are divided into multiple layers based on their importance to the prediction result: core features require high-precision decoding, while secondary features can accept approximate decoding. For scenarios with high real-time requirements, only the core feature layer may be decoded; for scenarios with high precision requirements, more layers of features will be decoded. This hierarchical architecture achieves a flexible balance between decoding precision and computational efficiency, adapting to different application needs.

[0371] A pre-computation and caching mechanism is implemented to accelerate the processing of common decoding patterns. For frequently occurring feature patterns or feature subsets, their decoding results or intermediate states are pre-computed and stored, and then directly retrieved from the table during the actual decoding process, avoiding redundant calculations. An efficient pre-computation strategy is designed: identifying high-frequency feature patterns (by analyzing historical data); generating decoding tables for these patterns; implementing a fast table lookup mechanism; and designing a cache replacement strategy (such as LRU, Least Recently Used) to manage limited cache space. This pre-computation and caching mechanism is particularly suitable for processing datasets with recurring patterns, and can significantly reduce the average decoding time.

[0372] By integrating the aforementioned technologies, a complete deterministic prediction model is constructed. This model comprises several core components: a data preprocessing module (standardizing input features and preparing for encoding); an FRS encoding module (implementing the encoding process in S702); a near-linear decoder (achieving efficient decoding); and a prediction layer (generating the final prediction based on the decoded features). An adaptation layer is also designed to automatically select the optimal decoding path based on the characteristics of the input data, balancing speed and accuracy. The entire model adopts a pipelined architecture, with components working in parallel to further improve throughput.

[0373] The deterministic prediction model underwent comprehensive testing and optimization to ensure its performance and stability. Testing included decoding accuracy evaluation (comparing with the theoretically optimal decoding results); time complexity verification (experimentally verifying near-linear time characteristics); memory usage analysis (ensuring suitability for the deployment environment); and performance curves for different data sizes (verifying the decoder's performance as the data size increases). Based on the test results, parameter optimization and performance tuning were performed, including FFT parameter settings, sparse recovery threshold adjustment, and cache size configuration.

[0374] The above design outputs a highly efficient deterministic prediction model with near-linear time complexity decoding capabilities. It significantly improves processing speed while maintaining prediction accuracy, making it particularly suitable for prediction scenarios requiring real-time response. The model is saved in a standard format, including encoding parameters, decoding algorithms, and prediction logic, facilitating deployment and integration into existing systems. This efficient prediction model provides strong technical support for the rapid assessment of project success rates.

[0375] Step S704: Based on the encoded feature data, design a randomized decoder with a full polynomial time complexity. The randomized decoder introduces a random sampling strategy to further reduce computational complexity while ensuring accuracy, and outputs a randomized prediction model.

[0376] In this sub-step, based on the encoded feature data, another type of efficient decoder—the randomized decoder—is designed. The randomized decoder complements the deterministic decoder from the previous step, providing flexible options for different application scenarios. This step embodies the core idea that "randomized decoding can be easily implemented using the Monte Carlo algorithm, resulting in a multinomial-time algorithm."

[0377] Full polynomial time complexity refers to an algorithm with a time complexity of O(n^c), where n is the input size and c is a constant. While the polynomial time complexity may theoretically be slightly higher than the near-linear time complexity of a deterministic decoder, randomization techniques can significantly reduce the computational constant in practice, making the algorithm more efficient in specific scenarios. A randomized algorithm is one that incorporates random selection during execution. Compared to deterministic algorithms, it may produce different results in different runs, but with proper design, the result can be guaranteed to be correct with a high probability, while significantly reducing computational complexity.

[0378] First, we analyze the theoretical basis of randomized decoding and clarify the design objectives. In coding theory, randomization techniques are often used to reduce decoding complexity, especially for complex error-correcting codes. Drawing on this idea, we design a randomized FRS decoder with the goal of reducing computational complexity by introducing randomness while maintaining decoding quality. Key design metrics are determined as follows: decoding accuracy (how close the decoded result is to the ideal result), computational complexity (theoretical and actual running time), error probability (the probability of decoding failure or inaccurate results), and parameter flexibility (whether the balance between accuracy and speed can be adjusted according to requirements).

[0379] A randomized decoding strategy based on the Monte Carlo method was designed. The Monte Carlo method is a class of computational algorithms based on random sampling, widely used to solve complex computational problems. In this paper, the Monte Carlo method is used to estimate the key components of the feature vector, rather than precisely calculating each component. The key steps to implement randomized decoding include: randomly selecting a subset of encoded features instead of using all encoded features; performing decoding computation based on this subset to obtain a preliminary estimate of the feature vector; obtaining multiple estimation results by repeating the above process multiple times; and combining these estimation results (e.g., by averaging or weighted averaging) to obtain the final decoding result. This method significantly reduces the computational cost of each decoding step, while ensuring the accuracy of the result through multiple sampling.

[0380] Importance sampling is employed to improve the efficiency of random sampling. Importance sampling is a variance reduction technique that adjusts the sampling distribution to give more important samples a higher sampling probability. In this study, the importance distribution of encoded features is analyzed to identify feature locations that have a greater impact on the decoding results. During the sampling process, these highly important features are prioritized while maintaining a certain proportion of randomness to balance exploration and utilization. This strategy significantly improves sampling efficiency, enabling the same number of samples to produce more accurate estimation results.

[0381] An adaptive sampling strategy was implemented to dynamically adjust sampling parameters. Adaptive sampling refers to a method that dynamically adjusts the sampling process based on the current estimation results and computational state. A multi-layered adaptive mechanism was designed: adaptive sample size (adjusting the number of samples based on the initial estimation confidence level); adaptive sampling distribution (adjusting the importance weights of subsequent samples based on the results of previous rounds of sampling); and adaptive stopping conditions (stopping sampling early when the estimation result reaches a preset accuracy or multiple estimation results tend to stabilize). This adaptive strategy enables the automatic finding of the optimal balance between accuracy and efficiency for decoding tasks of varying difficulty.

[0382] This paper incorporates the principles of the Las Vegas algorithm to ensure the accuracy of the decoding results. The Las Vegas algorithm is a special type of randomized algorithm characterized by a guaranteed correct result but randomized runtime. In this paper, the Las Vegas principle is manifested as follows: the correctness of the decoding result is verified (e.g., by re-encoding to check if it matches the original encoded features); if the result is found to be inaccurate, the number of samples is increased or a more precise decoding method is switched to; only when the result verification is successful is the final prediction output. This verification mechanism ensures the reliability of randomized decoding and prevents error accumulation and propagation.

[0383] A parallel and distributed computing architecture was designed to fully leverage the capabilities of modern computing platforms. A key advantage of randomized decoding is its inherent parallelizability; different sampling and decoding processes can be performed independently. A multi-layered parallel strategy was implemented: thread-level parallelism (multiple sampling tasks processed in parallel on a single machine with multiple cores); data-level parallelism (large feature vectors are divided into multiple blocks, processed in parallel, and the results are merged); and model-level parallelism (multiple decoders with different configurations run simultaneously, and the best result is selected). For large-scale applications, distributed computing is also supported, distributing decoding tasks to multiple servers and coordinating the computation process through a message passing interface (MPI) or a distributed computing framework (such as Spark).

[0384] By integrating the above technologies, a complete randomized prediction model is constructed. This model comprises several core components: a random sampling controller (determining the sampling strategy and parameters); a parallel decoding engine (executing multiple parallel decoding tasks); a result integrator (merging multiple decoding results); a quality checker (verifying the reliability of the decoding results); and a prediction generator (generating the final prediction based on decoding features). A dynamic configuration mechanism is also implemented, allowing model parameters to be adjusted according to application requirements, such as accuracy requirements, time constraints, and resource limitations.

[0385] A comprehensive evaluation and optimization of the randomized prediction model was conducted to verify its performance advantages. Testing included accuracy analysis (comparison with a deterministic decoder); speed testing (verifying the reduction in actual computation time); scaling performance (testing the model's performance on data of different scales); resource consumption (CPU and memory usage); and robustness testing (performance under different noise levels). Based on the test results, model parameters and architecture were optimized to find the optimal speed-accuracy balance.

[0386] The above design outputs a high-efficiency randomized prediction model with full multinomial time complexity. By introducing a random sampling strategy, computational complexity is further reduced while maintaining accuracy, making it particularly suitable for scenarios with extremely high speed requirements or limited computing resources. This model complements the deterministic prediction model, providing flexible options for different application scenarios. The model is saved in a standard format, including randomization parameters, decoding algorithms, and prediction logic, facilitating deployment and integration.

[0387] Step S705: Based on the deterministic prediction model and the randomized prediction model, a comprehensive prediction model is constructed using model ensemble techniques. The model parameters are optimized through cross-validation and grid search methods, and the project success rate prediction model is output.

[0388] In this sub-step, the deterministic and randomized prediction models built in the first two steps are integrated to construct a superior comprehensive prediction model. Model ensemble is a commonly used technique in machine learning to improve prediction performance; by combining the prediction results of multiple models, more accurate and stable predictions are obtained.

[0389] Model ensemble is a technique that combines multiple base models to build a more powerful model. The theoretical basis of ensemble learning is the principle of "diversity reduces variance," meaning that when multiple different models predict the same problem, their errors are often not entirely correlated. Appropriate combination can offset their individual weaknesses, producing a more accurate overall prediction. In this paper, deterministic and randomized prediction models have different algorithmic foundations and advantages: deterministic models provide stable and repeatable predictions, making them particularly suitable for scenarios requiring interpretability; randomized models, on the other hand, may have advantages in computational efficiency and handling specific types of data. By ensembled, these two models can combine their strengths and overcome their respective limitations.

[0390] First, we analyze the prediction patterns and error distributions of the two models to assess their complementarity. By comparing the prediction results of the two models on the validation set, we calculate several metrics: prediction consistency (the proportion of identical predictions by both models); error correlation (the correlation coefficient between the errors of the two models); and their respective advantage domains (the subset of data where each model performs best). These analytical results provide a basis for subsequent ensemble strategy design. Highly complementary models are suitable for ensembles with relatively balanced weights, while when one model significantly outperforms the other, it may be necessary to favor the stronger model.

[0391] Multiple ensemble strategies were implemented, and the optimal solution was selected through experiments. The simple averaging method is the most basic ensemble approach, directly taking the average of the predictions from the two models as the final prediction. The weighted averaging method assigns weights based on the overall performance of each model, with the better-performing model receiving a higher weight. The conditional weighting method is more flexible, dynamically adjusting weights based on the characteristics of the input features, using the most suitable model for different types of data. Stacking is a more complex ensemble approach, using another "meta-model" to learn how to best combine the predictions of the base models. A two-level stacking architecture was implemented: the first level consists of deterministic and randomized models running in parallel, generating base predictions; the second level is a meta-model (such as logistic regression, gradient boosting trees, or neural networks), which takes the predictions from the first-level model as input to generate the final prediction. This method can learn complex combination rules, surpassing simple linear combinations.

[0392] Cross-validation optimizes ensemble parameters to ensure the model's generalization ability. Cross-validation is a method for evaluating model performance and helping to tune hyperparameters. It reduces the risk of overfitting by dividing the data into multiple subsets and repeatedly training and validating the model. Using k-fold cross-validation (typically k=5 or 10), the training data is divided into k equal-sized subsets; each time, k-1 subsets are used to train the model, and the remaining subset is used for validation; this is repeated k times, with each subset used as a validation set; finally, the performance metrics from the k validations are summed to obtain the model's average performance. Cross-validation allows for a more reliable evaluation of the performance of different ensemble strategies, enabling the selection of the scheme with truly good generalization ability.

[0393] A grid search approach is employed to systematically explore the hyperparameter space and find the optimal configuration. Grid search is a hyperparameter optimization technique that determines the best parameter settings by trying all combinations in a predefined parameter grid. In this study, key parameters to be optimized include: ensemble weights (the relative importance of different models); stacked model parameters (if stacking is used); decoder parameters (e.g., FFT parameters for a deterministic decoder, sampling parameters for a randomized decoder); and feature selection thresholds (determining which features to use). A multidimensional parameter grid is constructed, and each parameter combination is cross-validated to select the combination with the best average performance. To improve efficiency, hierarchical grid search (a coarse search followed by a fine-grained search in promising regions) or random search (random sampling in the parameter space, suitable for high-dimensional parameter spaces) may be used.

[0394] Special emphasis is placed on model calibration and uncertainty estimation to provide reliable probabilistic predictions. Probabilistic calibration is the process of ensuring that the probability values ​​output by the model reflect the true uncertainty. The degree of model calibration is assessed using reliability diagrams and expected calibration error (ECE), and calibration performance is improved through techniques such as temperature scaling. Predictive uncertainty estimation is also implemented, providing confidence intervals or standard errors for each prediction. This is particularly important for predicting project success rates, as it helps decision-makers understand the reliability of the predictions. Uncertainty estimation methods include Bayesian methods (such as Bayesian neural networks), ensemble methods (utilizing the variance of predictions from multiple models), or direct modeling (training a model to simultaneously predict both mean and variance).

[0395] A comprehensive evaluation of the ensemble model is conducted to ensure its superiority across all performance metrics. Performance evaluation metrics include accuracy, precision, recall, F1 score (for binary classification problems, such as determining the success of an item); root mean square error (RMSE), mean absolute error (MAE) (for regression problems, such as predicting a specific success rate); ROC curve and area under the AUC (to assess the model's ability to distinguish between different classes); and log loss (to assess the quality of probability predictions). The model's performance on different data subsets is also evaluated to ensure it maintains good performance under various conditions.

[0396] Conducting feature importance analysis improves model interpretability. Interpretability refers to the ability to understand the model's decision-making process and influencing factors, which is particularly important for predicting project success rates. Multiple methods are applied to analyze feature importance: global feature importance (measuring the contribution of a feature to the overall prediction); local feature contribution (analyzing the contribution of each feature to a specific prediction instance); and partial dependency graphs (showing the relationship between a specific feature and the prediction result). These analyses not only help understand the model's decision-making basis but also provide specific guidance for improving project proposals.

[0397] In this embodiment, step S107 includes the following sub-steps:

[0398] Step S801: Based on the project success rate prediction model and the feature vector of the project to be evaluated, calculate the success rate of the project under different platforms and conditions, apply the deterministic decoder or the randomized decoder, select the appropriate decoding method according to the real-time requirements of the application scenario, and output the project success rate prediction result.

[0399] In this sub-step, the previously constructed prediction model is applied to predict the success rate of new projects to be evaluated. This is the practical application stage of the entire prediction process, requiring flexible selection of decoding methods based on actual needs to generate reliable prediction results.

[0400] First, feature extraction and processing are performed on the projects to be evaluated to obtain structured feature vectors. Key features are extracted from project application materials, company information, and external data sources, including basic project attributes (such as technology field, innovation type, and application amount), applicant company characteristics (such as size, R&D capabilities, and historical performance), team characteristics (such as personnel composition, educational background, and experience level), and industry characteristics (such as industry trends, policy guidance, and competitive landscape). These raw features undergo the same preprocessing steps as in the training phase, including data cleaning, missing value handling, standardization, and encoding transformation. For newly emerging feature values ​​(such as category values ​​not seen in the training set), specific processing strategies are adopted, such as mapping to the most similar known category or using encoding methods specifically designed for new values. The processed feature vectors maintain the same format and scale as the training data to ensure that the prediction model can interpret them correctly.

[0401] The system intelligently selects the appropriate decoding method based on the requirements of the application scenario. Different application scenarios may have varying requirements for real-time performance and accuracy. A scenario-adaptive mechanism is designed to automatically select the most suitable decoding method: when real-time response is required (e.g., online evaluation needs immediate results), the more computationally efficient randomized decoder is prioritized, potentially sacrificing a small amount of accuracy for speed; when high accuracy is required (e.g., final evaluation of important projects), the deterministic decoder is prioritized to ensure the reliability and stability of the prediction results; when resources are limited (e.g., mobile devices or heavily loaded servers), the decoding method with lower resource consumption is prioritized; when repeatability is required (e.g., providing the same evaluation results), the deterministic decoder must be selected. A hybrid strategy is also supported, such as first using a randomized decoder to quickly provide an initial evaluation, and then using a deterministic decoder for precise calculation as needed.

[0402] A comprehensive assessment of the success rate of projects under various conditions is conducted. Multi-condition assessment is an analytical method that considers the likelihood of project success under different scenarios. It considers multiple variables: different platform conditions (each possible review or funding platform); different time windows (near-term, mid-term, or long-term applications); different application strategies (e.g., emphasizing technological innovation, market application, or social benefits); and different levels of support (e.g., applying for different levels of support or funding amounts). The project success rate is calculated for each combination of conditions, generating multi-dimensional assessment results that comprehensively reflect the project's likelihood of success under different scenarios.

[0403] An ensemble decoder transforms encoded features into predicted outcomes. The decoding process comprises two main steps: feature decoding (mapping encoded features back to the original feature space or intermediate feature representation) and prediction generation (calculating success rates based on the decoded features). Depending on the previously selected decoding method, a deterministic decoder (using a near-linear-time algorithm for precise decoding) or a randomized decoder (using a multinomial-time algorithm based on random sampling for fast decoding) is applied. The decoded features are then transformed into specific success rate predictions through a prediction layer (such as logistic regression, random forest, or a neural network). For binary classification problems (success / failure), the success probability is output; for multi-class problems (such as different levels of success), the probability distribution for each class is output.

[0404] Alongside generating forecasts, the system provides uncertainty estimates and explanatory information. Uncertainty estimation, a method for quantifying forecast reliability, is particularly important for predicting project success rates. Calculating the confidence interval or standard error of the forecast helps understand the range of reliability for the results. Explanatory information is also provided, illustrating which factors have a positive or negative impact on the forecast, which factors are decisive, and how to improve the project to increase the success rate. These explanations are generated through feature importance analysis, local explanation models (such as LIME or SHAP), or rule-based explanation generators.

[0405] A quality assurance mechanism is implemented to ensure the reliability of the output. This includes range checks (ensuring predicted values ​​are within a reasonable range), consistency checks (verifying the relative relationships between predictions under different conditions), anomaly detection (identifying potentially inaccurate predictions), and manual review when necessary (predictions for important or boundary cases may be marked as requiring expert review). Key information about the prediction process, such as the model version used, decoding method, and feature quality assessment, is also recorded to facilitate subsequent auditing and improvement.

[0406] Through the above process, comprehensive project success rate prediction results are output, including success probabilities under different platforms and conditions, analysis of key influencing factors, estimation of prediction uncertainty, and improvement suggestions. These prediction results are both intuitive and easy to understand, and contain rich details, providing valuable reference for project applicants and decision-makers. A complete log of the prediction process is also recorded for subsequent model monitoring and improvement. The prediction results are output through a standard interface and can be seamlessly integrated into report generation or decision support.

[0407] Step S802: Based on the project success rate prediction results and historical support policy data, apply time series forecasting technology to analyze the future trend of support intensity changes, taking into account factors including policy cycles, economic environment and industry development, predict the changes in support intensity for different types of projects, and output the support intensity trend analysis results.

[0408] This sub-step involves predicting and analyzing the future level of policy support based on project success rate projections and historical support policy data. This trend prediction of support levels is crucial for strategic planning in project applications, helping applicants grasp the optimal application timing and strategies.

[0409] First, a comprehensive historical database of support policies is constructed as the foundation for trend prediction. This historical data includes policy documents (policy release date, content, priority areas, and support measures), funding data (total annual funding amount, average funding amount, and number of projects in each area), evaluation criteria (evaluation indicators and weight changes over different periods), and application requirements (application conditions, material requirements, and process changes). This raw data undergoes structuring processing, including time standardization (unifying time scales, such as quarterly or annual organization), area classification (establishing a unified area classification system to ensure data comparability across periods), indicator extraction (extracting quantitative indicators from policy texts, such as frequency of mention, priority statements, and funding allocation requirements), and correlation analysis (identifying the inheritance and development relationships between policies). The processed data is then organized in a time series for easier subsequent trend analysis.

[0410] A multi-layered time series analysis framework was designed to understand policy change patterns at different time scales. Time series analysis is a method that studies data points arranged in chronological order to extract meaningful statistical information and patterns. Three time scales of analysis were implemented: short-term analysis (quarterly to annual changes, capturing immediate adjustments), medium-term analysis (2-5 year cyclical changes, capturing the impact of planning cycles), and long-term analysis (trends over 5 years, capturing basic policy directions). Different analytical techniques were applied to each time scale: trend analysis (identifying long-term growth or decline trends in data, such as using moving averages or linear regression); seasonality analysis (identifying cyclical patterns, such as quarterly fund allocation patterns within a year); periodicity analysis (identifying fluctuations over longer periods, such as changes related to five-year plans or economic cycles); and irregular fluctuation analysis (identifying and interpreting non-cyclical changes, such as sudden policy adjustments).

[0411] A comprehensive support intensity prediction model is constructed by combining multiple factors. The factors influencing support policies are complex and diverse, and three core categories are considered: policy cycle factors (such as five-year planning cycles, annual budget cycles, and leadership transition cycles), economic environment factors (such as GDP growth rate, fiscal situation, and economic restructuring), and industry development factors (such as technological breakthroughs, industrial upgrading, international competition, and the rise of emerging fields). A factor-enhanced time series model is adopted, incorporating these external factors as predictive variables into the model. Specific implementations include ARIMAX models (autoregressive integrated moving average models combined with external variables, capable of handling the combined effects of linear time series and external factors), state-space models (such as dynamic linear models, capable of handling time-varying coefficients and multivariate inputs), or machine learning models (such as gradient boosting tree-based time series prediction, capable of capturing complex nonlinear relationships).

[0412] This system enables differentiated prediction of support levels based on project type. Different types of projects may experience varying support trends throughout the policy cycle. Based on dimensions such as industry sector, technological characteristics, innovation type, and company size, projects are categorized into multiple groups, and a dedicated prediction model is built for each group. The modeling process considers both the correlations (e.g., some groups may have similar policy sensitivities) and the differences (e.g., the policy cycles for traditional and emerging industries may differ significantly). The model training employs a hierarchical structure, including a shared base model (capturing general trends) and group-specific adjustment layers (capturing group characteristics).

[0413] Special attention is paid to predicting policy turning points, as these are the moments when changes in support intensity are most significant. Policy turning points refer to the points in time when policy direction or intensity changes markedly; accurately predicting these points is crucial for project application strategies. Changepoint detection technology is applied to identify policy turning points in historical data, analyze the characteristics and triggering factors before and after the turning points, and predict future possible turning points based on these patterns. Expert rules and external event calendars (such as important meetings and planning timelines) are also combined to improve the accuracy of turning point predictions. For predicted turning points, their probability of occurrence, potential direction of impact, and intensity are assessed to provide more refined references for decision-making.

[0414] This system generates multi-scenario forecasts of support levels to reflect future uncertainties. The future policy environment presents numerous possibilities, and a single forecast may not accurately reflect this uncertainty. Scenario analysis is employed to construct multiple possible future scenarios: a baseline scenario (continuation of current policy trends), a proactive scenario (increased support), a conservative scenario (weakened support), and a transitional scenario (changes in policy priorities). Each scenario corresponds to different assumptions and model parameter settings. Predictive results are generated for each scenario, and key factors for project success and strategic adjustment recommendations are analyzed under different scenarios. The probability of each scenario occurring is also assessed to help decision-makers understand the distribution of risks and opportunities.

[0415] Provide visualization and interpretation of the forecast results to ensure they are understandable and actionable. Visualizations include trend lines (showing the time-varying trend of support intensity), heatmaps (showing the support intensity for different project types at different times), radar charts (comparing policy priorities across different periods from multiple dimensions), and scenario comparison charts (comparing different forecast scenarios). Interpretations include key drivers of the trend (e.g., which policies or economic factors have the greatest impact on the trend), explanations of inflection points (predicting the reasons and signals of inflection points), opportunity and risk analysis (predicting the potential impact of the trend on projects), and operational suggestions (e.g., adjusting application timing, strengthening preparations in specific areas, optimizing project positioning, etc.).

[0416] The above analysis yields comprehensive trend analysis results on support intensity, including time-series forecasts of support intensity for different types of projects, forecasts of changes in policy priorities, forecasts of turning points, multi-scenario analysis, and strategic recommendations. These results provide both quantitative forecast data and qualitative trend interpretations, offering a scientific basis for strategic planning of project applications. The results are output in a standard format and can be directly used to generate forecast reports or provide decision support.

[0417] Step S803: Based on the project success rate prediction results and the support intensity trend analysis results, use data visualization technology to generate an intuitive prediction report. The prediction report includes the success rate value, confidence interval, key influencing factor analysis, and support intensity trend chart, and outputs the visualized prediction report.

[0418] This sub-step integrates the outputs of the first two steps—project success rate prediction results and support intensity trend analysis results—to generate a comprehensive, intuitive, and practical visual prediction report. Data visualization is a technique that transforms data into a visual representation, making complex analytical results easier to understand and use.

[0419] First, a multi-layered report structure was designed to ensure the logical flow and readability of the information. The report structure, as the framework for organizing and presenting information, directly impacts users' understanding and efficiency in using the content. A three-layer structure is adopted: a summary layer (providing key conclusions and readily understandable core indicators, such as overall success rate scores, optimal application platforms and timing, and key risk factors); an analysis layer (providing detailed forecast results and analysis, including multi-dimensional success rate analysis, support intensity trends, and key factor analysis); and a technical layer (providing methodological explanations, data sources, model parameters, and other technical details to meet the needs of professional users). This layered structure allows users with different needs to find the appropriate depth of information, benefiting everyone from senior decision-makers to technical analysts.

[0420] This application of data visualization principles and techniques transforms quantitative analysis results into intuitive graphical representations. Data visualization is a discipline that uses visual elements such as charts, graphs, and maps to represent data. It implements various visualization types: indicator dashboards (displaying core success rate indicators and health status, such as success rate scores, relative position assessments, and comparisons with industry averages); success rate heatmaps (using color depth to represent success rates under different platforms and conditions, facilitating quick identification of the best choice); trend line charts (showing the changing trend of support intensity over time, marking prediction turning points and critical periods); radar charts (displaying the project's strengths and weaknesses from multiple dimensions, such as scores for technological innovation, team strength, and business model); factor influence charts (such as waterfall charts or funnel charts, showing the positive and negative contributions of each factor to the success rate); decision tree visualizations (showing changes in success rate under different decision paths); and uncertainty visualizations (such as confidence interval bands and probability distribution charts, showing the range of uncertainty in predictions).

[0421] Special emphasis is placed on representing the reliability of success rate predictions to help users understand the degree of certainty in the prediction results. Uncertainty representation is a method to quantify and visualize the reliability of prediction results. Several uncertainty representation methods are designed: confidence interval bands (adding shaded areas to a trend line graph to represent the possible range of predicted values); probability density plots (showing the distribution of predicted success rates, not just a single value); confidence scores (providing a confidence score for each prediction, such as "high confidence," "moderate confidence," or "low confidence"); and sensitivity analysis (showing the impact of changes in key assumptions on the prediction results, helping to understand the robustness of the predictions). These representation methods help users correctly understand and use the prediction results, avoiding overconfidence or over-pessimism.

[0422] It provides visual analysis of key influencing factors to help users understand the drivers of project success and areas for improvement. Influence factor analysis is the process of identifying and quantifying the key variables affecting project success rates. Multiple visualization techniques are used to display influencing factors: factor importance ranking chart (a bar chart arranged by influence, visually showing the relative importance of each factor); factor contribution waterfall chart (showing the positive or negative contribution of each factor to the predicted value, and how it progresses from baseline to final prediction); factor interaction heatmap (showing the interaction effects between factors, such as the particularly strong impact of certain factor combinations); improvement simulator (an interactive tool that allows users to adjust specific factor values ​​and view the potential impact on success rates in real time, helping to plan improvement strategies); and benchmarking analysis (comparing the project with successful cases or industry benchmarks in key factors to identify gaps and areas for improvement).

[0423] This system generates dynamic visualizations of support intensity trends, showcasing anticipated changes in the policy environment. Trend visualization is a technique for displaying patterns in time-series data. It implements various trend visualization methods: multi-line trend charts (showing changing trends in support intensity for different types of projects or policy areas); seasonal heatmaps (showing seasonal patterns within a policy cycle, such as the intensity of funding disbursements in different months of the year); flow charts of policy focus changes (such as Sankey diagrams, showing the flow and changes in key policy areas over time); scenario comparison charts (showing changes in support intensity under different forecast scenarios side-by-side); event-marked timelines (marking important policy events and turning points in forecasts on the trend chart to help understand the driving factors behind trend formation); and forecast uncertainty funnel charts (showing that uncertainty increases the further back in the forecast period). These visualizations intuitively demonstrate the dynamic changes in the policy environment, helping applicants grasp opportunities and trends.

[0424] The design includes personalized recommendations and action suggestions, transforming the analysis results into actionable strategies. These recommendations and suggestions provide specific action guidance based on the analysis results. Multiple types of suggestions are generated: platform selection suggestions (recommending the most suitable application platform based on predicted success rates); timing selection suggestions (recommending the optimal application window based on support trends); project optimization suggestions (proposing specific directions for project improvement based on influencing factor analysis, such as strengthening certain technologies, adjusting team composition, and refining the business model); risk warnings (identifying key risk factors that may affect project success and strategies to mitigate these risks); and application strategy suggestions (such as key points for material preparation, direction for defense preparation, and resource allocation suggestions). These suggestions provide both quantitative support (such as the expected potential for increasing the success rate) and specific, actionable steps, helping applicants translate the analysis results into concrete actions.

[0425] Interactive features have been implemented in the report, enhancing user experience and information retrieval efficiency. Interactive visualizations allow users to actively explore data, rather than passively receiving information. Multiple interactive functions are provided: dynamic filtering (allowing users to filter data based on specific conditions, such as viewing only forecast results for a specific industry or company size); drill-down (allowing users to access more detailed analysis levels from a high-level overview); parameter adjustment (allowing users to modify certain assumptions and see the impact on forecast results); customizable views (allowing users to select indicators and charts of interest); annotations and markers (allowing users to add their own annotations and highlighting elements to the report); and report export (supporting exporting reports as PDF, PowerPoint, or interactive web pages). These interactive features transform the report from a static document into a dynamic decision-making tool, significantly increasing its practical value.

[0426] like Figure 3 As shown, this embodiment of the invention also provides an intelligent prediction device for project success rate, comprising:

[0427] The data acquisition module 901 is used to acquire multi-source data related to the project. Through data cleaning and standardization, named entity recognition and relation extraction, a structured knowledge base containing successful project cases and their characteristics is constructed.

[0428] The feature extraction module 902 is used to extract features and search substructures from the JSONL format project data in the structured knowledge base by applying the merging tree representation technique to preserve the identity of individual projects while constructing a unified tree structure representation. It also constructs a data structure through an extended Burrows-Wheeler transformation and combines path decomposition, ancestor calculation and adaptive tree identifier collection techniques to obtain a project feature set.

[0429] The multidimensional analysis module 903 is used to analyze the core dynamics of project development from the time dimension and the fundamental characteristics of the project enterprise from the spatial dimension based on the project feature set, and to generate a multidimensional feature vector through a feature fusion algorithm.

[0430] The allocation model construction module 904 is used to divide projects into different groups based on the multidimensional feature vector, construct a project-platform allocation model with group fairness, and model the overall platform constraints and group-specific constraints through a convex cost function to realize a soft constraint mechanism of restrictive dominance and minority group protection.

[0431] The optimization solution module 905 is used to design a polynomial-time approximation algorithm based on linear programming and network flow to solve for the optimal allocation scheme for the project-platform allocation model.

[0432] The prediction model building module 906 is used to perform feature encoding using improved folded Reed-Solomon code technology based on the optimal allocation scheme and historical project success data, and to build a project success rate prediction model through deterministic and randomized decoders.

[0433] The result generation module 907 is used to assess the success rate of the project to be evaluated and predict the trend of support intensity based on the project success rate prediction model, and generate a visualized prediction report.

[0434] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent prediction of project success rate, characterized in that, Includes the following steps: Acquire multi-source data related to the project, and construct a structured knowledge base containing successful project cases and their characteristics through data cleaning and standardization, named entity recognition, and relation extraction. For the JSONL formatted project data in the structured knowledge base, a merging tree representation technique is applied to preserve the identity of individual projects while constructing a unified tree structure representation. An extended Burrows-Wheeler transformation is used to construct the data structure. Combined with path decomposition, ancestor calculation, and adaptive tree identifier collection techniques, feature extraction and substructure search are achieved to obtain a project feature set. This includes: for the JSONL formatted project data in the structured knowledge base, a JSON parser is applied to parse each project data into an independent tree structure, preserving the hierarchical relationship of key-value pairs, assigning a unique identifier to each tree node to ensure that the original project can be traced in subsequent processing, and outputting a set of project tree structures; based on the set of project tree structures, a tree merging algorithm is applied to merge the tree structures of multiple projects into a unified merged tree representation. During the merging process, the independent identity information of each project is preserved, and different projects are distinguished by introducing identifier nodes and path prefixes. For the same path, output a merged tree representation; apply the extended Burrows-Wheeler transformation algorithm to the merged tree representation to transform the tree structure into an index data structure containing three parts: a node label array, a tree topology array, and a project identifier array, supporting fast tree navigation and sub-path query, and output an extended Burrows-Wheeler transformation index structure; based on the extended Burrows-Wheeler transformation index structure, combined with path decomposition, ancestor calculation, and adaptive tree identifier collection technology, achieve efficient substructure search of project data, extract key features including technical field, innovation points, team size, and funding needs from project data through predefined feature patterns, and output a project feature set; wherein, the ancestor calculation is the process of determining the ancestor-descendant relationship between two nodes, and the adaptive tree identifier collection technology can automatically adjust the identifier collection strategy according to query requirements; Based on the project feature set, the core dynamics of project development are analyzed from the time dimension, and the fundamental characteristics of the project enterprise are analyzed from the spatial dimension. Through feature fusion algorithm, a multi-dimensional feature vector is generated. Based on the multidimensional feature vector, the project is divided into different groups, and a project-platform allocation model with group fairness is constructed. The overall platform constraints and group-specific constraints are modeled through a convex cost function to realize a soft constraint mechanism of restrictive dominance and minority group protection. For the project-platform allocation model, a multinomial-time approximation algorithm based on linear programming and network flow is designed to solve for the optimal allocation scheme. The optimal allocation scheme includes the allocation decision for each project, the cost and utility evaluation of the scheme, the constraint satisfaction, and the approximation ratio guarantee. Based on the optimal allocation scheme and historical project success data, feature encoding is performed using the improved folded Reed-Solomon code technique to map the high-dimensional feature space to a low-dimensional encoding space. A project success rate prediction model is then constructed using a deterministic decoder and a randomized decoder. Based on the project success rate prediction model, the success rate and support trend of the projects to be evaluated are predicted, and the project success rate prediction results and support trend analysis results are output respectively, forming a visualized prediction report; wherein, the project success rate prediction results include the success probability under different platforms and conditions, key influencing factor analysis, prediction uncertainty estimation and improvement suggestions.

2. The method according to claim 1, characterized in that, The process involves acquiring multi-source data related to the project, and through data cleaning and standardization, named entity recognition, and relation extraction, constructing a structured knowledge base containing successful project cases and their characteristics, including: Based on web crawling technology and API interfaces, we collect multi-source data including lists of successful projects, guide documents, interpretation documents, and related news. We use regular expressions and text segmentation technology to perform preliminary processing on the collected unstructured data, remove irrelevant information, standardize the data format, and output the preliminary processed dataset. The dataset that has been initially processed is cleaned using natural language processing techniques, including removing duplicate content, correcting typos, and standardizing the format. At the same time, named entity recognition technology is used to extract key entity information and output a standardized entity dataset. Based on the standardized entity dataset, a relation extraction algorithm is applied to identify the relationships between entities, and knowledge graph technology is used to represent entities and relationships in a structured manner to construct a knowledge graph of successful project cases and output an initial knowledge base. The initial knowledge base is optimized using knowledge reasoning and redundancy elimination algorithms. Distributed storage technology and indexing mechanisms are employed to store the optimized initial knowledge base as structured data in JSONL format, and the structured knowledge base is then output.

3. The method according to claim 1, characterized in that, Based on the project feature set, the core dynamics of project development are analyzed from a time dimension, and the fundamental characteristics of the project enterprise are analyzed from a spatial dimension. A feature fusion algorithm is used to generate a multi-dimensional feature vector, including: Based on the project feature set, time attributes including project application time, approval time, and execution cycle are extracted. Time series analysis algorithms are applied to calculate the time distribution, seasonal fluctuations, and long-term trends of the number of projects, as well as the time patterns of changes in regulatory policies, and output time dimension feature vectors. Based on the project feature set, the spatial attributes of the geographical distribution, industry distribution, and scale distribution of the enterprises to which the projects belong are extracted. Spatial clustering and distribution analysis algorithms are applied to calculate the differences in project success rates in different regions and industries, as well as the correlation between enterprise scale and project success, and output spatial dimension feature vectors. Based on the project feature set, we extract strength indicators including the company's asset size, R&D investment, team composition, and historical performance. We apply feature engineering techniques to calculate the comprehensive score and hierarchical classification of the company's strength, analyze the correspondence between the company's strength and project success, and output the company's strength feature vector. For the time-dimensional feature vector, spatial-dimensional feature vector, and enterprise strength feature vector, a feature fusion algorithm is applied to weight and fuse the features of the time-dimensional, spatial-dimensional, and enterprise strength dimensions, taking into account the correlation and importance between features of different dimensions, to generate the multi-dimensional feature representation.

4. The method according to claim 1, characterized in that, Based on the multidimensional feature vectors, projects are divided into different groups, and a project-platform allocation model with group fairness is constructed. A convex cost function is used to model the overall platform constraints and group-specific constraints, realizing a soft constraint mechanism of restrictive dominance and minority group protection, including: Based on the multidimensional feature vector, a clustering algorithm is applied to divide the projects into different groups according to dimensions including industry category, technology field, and enterprise size. A weight coefficient is assigned to each group, and the project group division results are output. Based on historical platform statistics and fair allocation standards, the service capacity limit of each platform and the acceptance ratio requirements of each group are determined. Convex function modeling technology is applied to construct a convex cost function of the overall platform constraint, so that when the total number of projects allocated to the platform exceeds the soft constraint, a non-linearly increasing cost penalty is generated, and the platform constraint model is output. Based on the project group division results, fairness constraint modeling technology is applied to construct a specific convex cost function for each platform and each group. This convex cost function promotes group fairness by penalizing unbalanced allocation, realizes a soft constraint mechanism of restrictive dominance and minority group protection, and outputs a group fairness constraint model. Based on the matching degree assessment of projects and platforms, a utility value is defined for each project-platform pair, an overall utility function is constructed, and a comprehensive optimization objective function is constructed by combining the platform constraint model and the group fairness constraint model.

5. The method according to claim 1, characterized in that, The aforementioned project-platform allocation model employs a polynomial-time approximation algorithm based on linear programming and network flow to solve for the optimal allocation scheme, including: Based on the project-platform allocation model, a linear programming relaxation problem is constructed. The convex cost function in the original problem is transformed into a set of linear constraints through piecewise linearization, while retaining the utility constraints. This forms a relaxation problem that can be solved using standard linear programming methods, and the linear programming model is output. The linear programming model is solved using the simplex method or the interior point method. Fractional solutions are obtained during the solution process, and the optimal solution of the linear programming problem is output. Based on the optimal solution of the linear programming, a corresponding network flow model is constructed, with the project as the source and the platform as the sink. The score allocation in the optimal solution of the linear programming is converted into the flow in the network. The minimum cost flow problem that satisfies capacity constraints and flow conservation is constructed, and the network flow model is output. For the network flow model, the minimum cost flow algorithm is applied to obtain the integer flow solution, thereby determining the integer allocation scheme from the project to the platform. Theoretical analysis ensures that the cost difference between the approximate solution and the global optimal solution does not exceed the preset approximation ratio, while satisfying the utility constraint, and outputs the optimal allocation scheme.

6. The method according to claim 1, characterized in that, The process involves using an improved folded Reed-Solomon code technique for feature encoding based on the optimal allocation scheme and historical project success data. This maps the high-dimensional feature space to a low-dimensional encoding space. A project success rate prediction model is then constructed using a deterministic decoder and a randomized decoder, including: Based on the optimal allocation scheme and historical project success data, a model training dataset is constructed. Through data cleaning and standardization, including feature standardization, missing value handling, and outlier detection, the processed training dataset is output. The processed training dataset is then used to perform feature encoding using an improved folded Reed-Solomon code technique, which maps the high-dimensional feature space to a low-dimensional encoding space and outputs the encoded feature data.

7. The method according to claim 6, characterized in that, The step of using the optimal allocation scheme and historical project success data to perform feature encoding using improved folded Reed-Solomon codes, mapping the high-dimensional feature space to a low-dimensional encoding space, and constructing a project success rate prediction model through a deterministic decoder and a randomized decoder, further includes: Based on the encoded feature data, a deterministic decoder with near-linear time complexity is designed. The deterministic decoder can complete the decoding process in near-linear time and output a deterministic prediction model. Based on the encoded feature data, a randomized decoder with a full polynomial time complexity is designed. The randomized decoder introduces a random sampling strategy to further reduce computational complexity while ensuring accuracy, and outputs a randomized prediction model. Based on the deterministic prediction model and the randomized prediction model, a comprehensive prediction model is constructed using model ensemble techniques. The model parameters are optimized through cross-validation and grid search methods, and the project success rate prediction model is output.

8. The method according to claim 1, characterized in that, Based on the project success rate prediction model, the success rate and support intensity trend of the project to be evaluated are predicted, and the project success rate prediction results and support intensity trend analysis results are output respectively, forming a visualized prediction report, including: Based on the project success rate prediction model and the feature vector of the project to be evaluated, the success rate of the project is calculated under different platforms and conditions. The deterministic decoder or the randomized decoder is applied, and the appropriate decoding method is selected according to the real-time requirements of the application scenario to output the project success rate prediction result. Based on the project success rate prediction results and historical support policy data, time series forecasting technology is applied to analyze the future trend of support intensity. Taking into account factors including policy cycles, economic environment and industry development, the support intensity changes of different types of projects are predicted, and the support intensity trend analysis results are output. Based on the project success rate prediction results and the support intensity trend analysis results, data visualization technology is used to generate an intuitive prediction report. The prediction report includes the success rate value, confidence interval, key influencing factor analysis, and support intensity trend chart, and the visualized prediction report is output.

9. A project success rate intelligent prediction device, characterized in that, include: The data acquisition module is used to acquire multi-source data related to the project. Through data cleaning and standardization, named entity recognition and relation extraction, a structured knowledge base containing successful project cases and their characteristics is constructed. The feature extraction module is used to extract features from the JSONL format project data in the structured knowledge base. It applies a merging tree representation technique to preserve the identity of individual projects while constructing a unified tree structure representation. The module uses an extended Burrows-Wheeler transformation to build the data structure, and combines path decomposition, ancestor calculation, and adaptive tree identifier collection techniques to achieve feature extraction and substructure search, resulting in a project feature set. This set includes: parsing each project data point into an independent tree structure using a JSON parser, preserving the hierarchical relationship of key-value pairs, assigning a unique identifier to each tree node to ensure traceability to the original project in subsequent processing, and outputting a set of project tree structures; and merging the tree structures of multiple projects into a unified merged tree representation using a tree merging algorithm, preserving the independent identity information of each project during the merging process, and distinguishing them by introducing identifier nodes and path prefixes. For identical paths in different projects, a merged tree representation is output. This merged tree representation is then processed using an extended Burrows-Wheeler transformation algorithm to convert the tree structure into an index data structure containing three parts: a node label array, a tree topology array, and a project identifier array. This supports fast tree navigation and sub-path queries, outputting an extended Burrows-Wheeler transformation index structure. Based on this extended Burrows-Wheeler transformation index structure, combined with path decomposition, ancestor calculation, and adaptive tree identifier collection techniques, efficient substructure search of project data is achieved. Key features, including technical fields, innovation points, team size, and funding needs, are extracted from the project data using predefined feature patterns, outputting a project feature set. The ancestor calculation process determines the ancestor-descendant relationship between two nodes, and the adaptive tree identifier collection technique automatically adjusts the identifier collection strategy according to query requirements. The multidimensional analysis module is used to analyze the core dynamics of project development from the time dimension and the fundamental characteristics of the project enterprise from the spatial dimension based on the project feature set, and to generate multidimensional feature vectors through feature fusion algorithms. The allocation model construction module is used to divide projects into different groups based on the multidimensional feature vector, construct a project-platform allocation model with group fairness, and model the overall platform constraints and group-specific constraints through a convex cost function to realize a soft constraint mechanism of restrictive dominance and minority group protection. The optimization solution module is used to design a multinomial-time approximation algorithm based on linear programming and network flow for the project-platform allocation model to solve for the optimal allocation scheme. The optimal allocation scheme includes the allocation decision for each project, the cost and utility evaluation of the scheme, the constraint satisfaction, and the approximation ratio guarantee. The prediction model building module is used to perform feature encoding using the improved folded Reed-Solomon code technique based on the optimal allocation scheme and historical project success data, mapping the high-dimensional feature space to a low-dimensional encoding space, and building a project success rate prediction model through a deterministic decoder and a randomized decoder. The results generation module is used to predict the success rate and support trend of the project to be evaluated based on the project success rate prediction model, and output the project success rate prediction results and support trend analysis results respectively to form a visualized prediction report; wherein, the project success rate prediction results include the success probability under different platforms and conditions, key influencing factor analysis, prediction uncertainty estimation and improvement suggestions.

Images