Enterprise project information prediction method and system based on machine learning

Abstract

The invention relates to an enterprise project information prediction method and system based on machine learning, and the method comprises the following steps: collecting enterprise data, and processing the enterprise data to obtain preprocessed data; extracting multi-dimensional enterprise features to construct enterprise feature vectors based on the preprocessed data, and dividing enterprises into a plurality of categories based on the enterprise feature vectors; aiming at each category obtained by division, respectively constructing and training an independent project information prediction model; performing intra-class verification and cross-class verification on the project information prediction model of each class, and dynamically adjusting classification of the classes or selection of the enterprise features according to a verification result; and inputting the data of the target enterprise into the corresponding project information prediction model, and outputting a project information prediction result. According to the method, a full-link closed loop from multi-source heterogeneous data integration to multi-modal accurate prediction to interpretable decision support is realized, and a high-accuracy, self-adaptive and trusted intelligent solution is provided for enterprise project information analysis.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence application technology, and in particular to a method and system for predicting enterprise project intelligence based on machine learning. Background Technology

[0002] With the rapid development of big data and artificial intelligence technologies, corporate behavior data is becoming increasingly abundant, covering multiple dimensions such as business registration information, financial statements, intellectual property, recruitment dynamics, policy applications, land and electricity use, and supply chain data. In scenarios such as government investment promotion, industrial investment, and corporate strategic decision-making, the demand for accurate prediction of corporate project development direction, project needs, regional site selection, and resource allocation is becoming increasingly urgent, and there is a pressing need for an efficient corporate project intelligence prediction technology to support decision-making.

[0003] Currently, traditional enterprise analysis methods mainly rely on manual surveys, expert experience, or simple statistical models. In existing technologies, data sources are scattered and lack a unified integration mechanism, making it difficult to break down data silos; the analysis dimensions are single, failing to comprehensively depict the enterprise's development intentions; the prediction models are mostly static, lacking adaptive learning and dynamic optimization capabilities, resulting in low prediction accuracy and an inability to effectively identify the enterprise's future project direction and site selection preferences; at the same time, existing technologies lack interpretable outputs, making it difficult to provide actionable decision support for government investment promotion, park operation, and industrial investment. These shortcomings are the core problems that this invention aims to solve. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a machine learning-based enterprise project intelligence prediction method and system. This system achieves a closed-loop process from multi-source heterogeneous data integration to multi-modal accurate prediction and interpretable decision support, providing a highly accurate, adaptive, and reliable intelligent solution for enterprise project intelligence analysis.

[0005] The above-mentioned objective of this invention is achieved through the following technical solutions: A machine learning-based method for predicting enterprise project intelligence includes the following steps: Collect enterprise data and process the enterprise data to obtain preprocessed data; Based on the preprocessed data, multi-dimensional enterprise features are extracted to construct enterprise feature vectors, and enterprises are divided into multiple categories based on the enterprise feature vectors. For each of the defined categories, an independent project intelligence prediction model is constructed and trained. For each category of project intelligence prediction model, intra-category and cross-category validations are performed, and the category division or the selection of enterprise characteristics is dynamically adjusted based on the validation results. Input the target company's data into the corresponding project intelligence prediction model, and output the project intelligence prediction result.

[0006] The above technical solutions construct a complete technical chain of "data processing, feature extraction and classification, cluster modeling, verification and adjustment, and prediction output", which solves the problems of scattered data sources, single analysis dimensions and low prediction accuracy of traditional methods. It enables systematic and accurate prediction of enterprise project intelligence and provides scientific decision-making basis for government investment promotion, industrial investment and other scenarios.

[0007] As a further technical solution of the present invention: the process of collecting enterprise data and processing the data to obtain preprocessed data specifically includes: Collect raw enterprise data from multiple pre-defined heterogeneous data sources; The original data of the enterprise is preprocessed; The preprocessing includes format normalization, missing value handling, error data correction, outlier handling, and redundant information removal.

[0008] By integrating the original enterprise data from multiple heterogeneous data sources through the above technical solutions, and through multi-stage preprocessing such as missing value handling, error data correction, outlier handling, and redundant information removal, data silos are broken down, data quality is improved, and clean, unified, and high-quality data input is provided for subsequent feature extraction and model building, avoiding interference from poor-quality data on prediction results.

[0009] As a further technical solution of the present invention: based on the preprocessed data, multi-dimensional enterprise features are extracted and an enterprise feature vector is constructed, specifically including: From the preprocessed data, time series features, spatial features, text features, and graph structure features are extracted; The extracted features are processed through feature selection, feature construction, feature encoding, and feature scaling to build enterprise feature vectors and form an enterprise feature library.

[0010] Through the above technical solutions, features of multiple types of enterprises are extracted and systematically processed to construct feature vectors, forming a structured enterprise feature library. This comprehensively portrays a multi-dimensional profile of enterprises, solving the shortcomings of traditional analysis with its single dimension. It provides rich and reusable feature support for enterprise classification and model training, and solidifies the foundation for prediction accuracy.

[0011] As a further technical solution of the present invention: the step of classifying enterprises into multiple categories specifically includes: Based on the enterprise feature library, and integrating multiple preset business dimensions, an industry-space-lifecycle ternary embedding is constructed as an enterprise feature representation. An adaptive density peak clustering algorithm is used to automatically divide enterprises into multiple homogeneous subgroups based on the enterprise feature representation.

[0012] Through the above technical solution, based on the feature library, six core business dimensions are integrated, including industry, position in the industrial chain, scale, region, life cycle, and equity structure. A three-element embedded feature of "industry-space-life cycle" is constructed. Combined with the adaptive density peak clustering algorithm, accurate classification of enterprises is achieved, ensuring that similar enterprises have similar features in the core business dimensions and dissimilar enterprises have significant differences. This provides a highly adaptable subset of enterprises for subsequent cluster modeling and effectively avoids the prediction bias caused by traditional "one-size-fits-all" modeling.

[0013] As a further technical solution of the present invention: for each of the categories, an independent project intelligence prediction model is constructed and trained, specifically as follows: An ensemble prediction model is constructed for each category, wherein the ensemble prediction model is configured to fuse the outputs of multiple sub-models that process multiple data types.

[0014] Through the above technical solutions, an independent integrated prediction model is built for each enterprise category, realizing a personalized modeling logic of "one model per category". This adapts to the characteristic differences of different enterprise categories, solves the problem of poor adaptability of traditional single models, and improves the predictive relevance and accuracy of the model for project intelligence of specific enterprise categories.

[0015] As a further technical solution of the present invention: the integrated prediction model integrates the outputs of the following sub-models through Bayesian fusion: LightGBM or XGBoost models are used for processing structured data; BERT model for processing text data; The GraphSAGE model is used for processing spectral data. And the Transformer model for processing time-series data.

[0016] By employing the above technical solutions, a Bayesian fusion approach is used to integrate and adapt sub-model outputs that are compatible with different data types. This fully leverages the advantages of each algorithm in processing structured, text, graph, and time-series data, enabling collaborative modeling of multimodal data and further improving the accuracy and robustness of prediction results for project type, location, and demand.

[0017] As a further technical solution of the present invention: the step of dynamically adjusting the category division or the selection of enterprise characteristics based on the verification results includes: If the performance evaluation index of the project intelligence prediction model in cross-class verification is higher than the preset threshold, it will trigger the reclassification of the category or the re-screening of the enterprise characteristics.

[0018] On the other hand, the present invention also discloses an enterprise project intelligence prediction system based on machine learning, including: a data acquisition and preprocessing module, used to acquire enterprise-related data from multiple heterogeneous data sources and preprocess the acquired data to obtain preprocessed data; The feature extraction and classification module is used to extract multi-dimensional enterprise features from the preprocessed data to construct enterprise feature vectors, and to classify enterprises into multiple categories based on the enterprise feature vectors. The cluster modeling module is used to build and train independent project intelligence prediction models for each of the categories. The verification and optimization module is used to perform intra-class and cross-class verification on the project intelligence prediction model for each category, and to trigger adjustments to the category division in the feature extraction and classification module or to trigger adjustments to the feature selection in the feature extraction and classification module based on the results of the cross-class verification. The prediction output module is used to input the target company's data into the corresponding project intelligence prediction model and output the project intelligence prediction result.

[0019] The above technical solution breaks down the prediction method into five functionally distinct and collaborative modules, enabling the engineering implementation of the technical solution. This ensures efficient connection between data collection, feature processing, modeling verification, and prediction output, providing users with a stable and operable enterprise project intelligence prediction system that meets the practical application needs of scenarios such as government investment promotion and park operation.

[0020] On the other hand, the present invention also discloses a computer device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the above-mentioned machine learning-based enterprise project intelligence prediction method.

[0021] The above technical solutions enable machine learning-based enterprise project intelligence prediction methods to run on computer devices. With the help of processors and memory, efficient execution of data processing, model training, and prediction output is achieved, providing hardware support for the large-scale application and promotion of the technical solutions.

[0022] On the other hand, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned machine learning-based enterprise project intelligence prediction method.

[0023] The above technical solution stores the prediction method in the form of a computer program on a readable medium, which facilitates the dissemination, deployment and reuse of the technical solution, lowers the threshold for users to obtain and apply the prediction method, and helps the technical solution to be widely implemented in different scenarios and among different user groups.

[0024] In summary, the present invention has at least one of the following beneficial technical effects: 1. This invention discloses a machine learning-based enterprise project intelligence prediction method and system. It adopts a complete chain of "multi-source data integration and preprocessing - multi-dimensional feature extraction and clustering - cluster modeling and fusion - dual verification and dynamic adjustment". It combines five functional modules, namely data acquisition and preprocessing, feature extraction and classification, cluster modeling, verification and optimization, and prediction output, to accurately predict enterprise project direction, site selection and resource requirements. It solves the pain points of traditional methods such as scattered data, low accuracy and lack of dynamic optimization. It also realizes the engineering implementation of technical solutions and provides interpretable and implementable decision support for government investment promotion, park operation and other scenarios.

[0025] 2. This invention discloses a device and a computer-readable storage medium for predicting enterprise project intelligence based on machine learning. It is supported by the processor and memory hardware of the computer device to ensure the efficient execution and large-scale operation of the prediction method, providing a stable hardware carrier for the practical application of the technical solution. At the same time, with the help of the computer-readable storage medium, the prediction method can be conveniently stored, disseminated and deployed in the form of computer programs, reducing the threshold for users to obtain and apply it, and helping the prediction technology to be widely implemented in different scenarios and different user groups. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the process of an enterprise project intelligence prediction method based on machine learning according to the present invention.

[0027] Figure 2 for Figure 1 A flowchart of S1.

[0028] Figure 3 for Figure 1 A flowchart of the S2 process.

[0029] Figure 4 for Figure 1 A flowchart of the S3 process.

[0030] Figure 5 for Figure 1 A flowchart of the S4 process.

[0031] Figure 6 for Figure 1 A flowchart of the S5 process.

[0032] Figure 7This is an architecture diagram of an enterprise project intelligence prediction system based on machine learning, according to the present invention. Detailed Implementation

[0033] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0034] In the description of this application, it should be noted that the terms "upper," "lower," "inner," "outer," "top / bottom," etc., indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0035] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installed," "equipped with," "sleeved / connected," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. Example

[0036] Reference Figure 1 This invention provides a machine learning-based enterprise project intelligence prediction method. This method follows an ordered logic of "data processing → feature construction → cluster modeling → validation and optimization → prediction output," specifically including five core steps executed sequentially: S1 data collection and preprocessing, S2 feature extraction and enterprise classification, S3 cluster modeling, S4 model validation and dynamic adjustment, and S5 project intelligence prediction and decision support. Each step is closely connected, with a clear data flow, collectively forming a complete enterprise project intelligence prediction chain. The method includes: Reference Figure 2 S1 data acquisition and preprocessing, in order to obtain comprehensive, timely and high-quality enterprise basic data, requires the completion of multi-source data acquisition and standardized preprocessing, which is specifically divided into two stages: S11 Data Acquisition: To break down data silos and ensure data coverage across all dimensions of business operations, raw information needs to be collected from multiple heterogeneous data sources. Specifically, raw data from enterprises is collected from various pre-defined heterogeneous data sources, such as business registration information, financial statements, intellectual property, recruitment dynamics, policy applications, land and electricity use, and supply chain data. The comprehensiveness and timeliness of data collection are ensured through technologies such as API calls, web crawlers, and direct database connections. A regular data update process is established to maintain information freshness and ensure that subsequent modeling data can reflect the latest business status of the enterprise.

[0037] S12 Data Preprocessing: To eliminate format differences, missing data, errors, anomalies, and redundancy in the raw data, a systematic cleaning operation is performed on the collected enterprise raw data. The core cleaning methods and specific implementation are as follows: Format standardization: Conduct format standardization processing to unify the formats of key fields such as date format, currency unit, and encoding format, ensuring that the format of all data is consistent, laying the foundation for subsequent feature extraction; Data quality optimization includes missing value handling and erroneous data correction. Missing value handling uses interpolation (to adapt to continuously missing time-series data), mean imputation (to adapt to evenly distributed numerical missing fields), or machine learning algorithm prediction imputation (to predict missing values ​​based on relevant features by training a simple regression model) to complete missing data and ensure data integrity. Erroneous data correction uses logical verification and domain knowledge to identify and correct logical contradictions, data entry errors, and other problems in the data, thereby improving data accuracy. Outlier handling: Statistical methods such as Z-score and IQR or box plot visualization tools are used to detect data anomalies, analyze the causes of outliers and determine their validity, and handle invalid outliers by deletion, replacement or data smoothing to avoid poor data interfering with subsequent processes; Redundant information removal: Delete duplicate fields or invalid information without differentiation in the data, reduce the data volume, and improve the efficiency of subsequent data processing.

[0038] Through the above cleaning operations, high-quality standardized data is finally output, providing reliable data input for the subsequent feature extraction process.

[0039] Reference Figure 3 S2, Feature Extraction, Feature Vector Construction, and Enterprise Classification: To comprehensively depict enterprise profiles and achieve accurate grouping, this step is broken down into three sub-processes: "S21 Feature Extraction → S22 Feature Vector Construction → S23 Enterprise Classification," as detailed below: Reference Figure 3Feature extraction in S21, as the core starting point of the S2 process, aims to extract high-value enterprise features from the high-quality preprocessed data output from S12, laying a solid foundation for subsequent feature vector construction and accurate enterprise classification. This step ensures feature quality and coverage through a combination of "pre-optimization + multi-dimensional extraction," and its implementation consists of two sub-steps: S211 Feature Extraction Preprocessing: Building upon the global data cleaning in S12, this section focuses on supplementary optimization of feature dimensions to improve feature quality. It calibrates feature field naming conventions and data type subdivisions to ensure consistent formatting at the feature level; it specifically detects hidden anomalies such as single feature distribution shifts and local extreme values ​​to avoid interference from invalid features; and it removes highly correlated, repetitive, or indistinguishable fields between features, simplifying dimensions while retaining high-value information, thus laying a solid foundation for multi-dimensional feature extraction.

[0040] S212 multi-dimensional feature extraction extracts four core original features, comprehensively covering the dimensions of enterprise dynamic trends, semantic intent, and external connections. Specifically, these include: time series features, extracted from dynamic data such as land use data, electricity consumption data, and revenue changes, to reflect the trend patterns of enterprise operation and development; spatial features, extracted from geographically relevant data such as the enterprise's registered location, business coverage area, and industrial chain spatial layout, to reflect the spatial distribution attributes of the enterprise; textual features, extracted from unstructured data such as recruitment dynamics, policy application documents, and enterprise announcements, to uncover the enterprise's development intent and business direction; and graph structure features, extracted after constructing an enterprise network based on supply chain relationships and equity association data, to reflect the enterprise's external connection attributes.

[0041] Reference Figure 3 S22 Feature Vector Construction and Enterprise Feature Library Establishment: Building upon the achievements of S21 Multi-dimensional Feature Extraction, the core is to generate usable feature vectors through systematic feature engineering optimization and to build a structured enterprise feature library. This provides standardized and reusable feature support for subsequent S23 accurate enterprise classification and S3 cluster modeling. The specific implementation consists of three sub-steps: S221 Feature Engineering Processing: Perform a series of optimization operations on the four types of original features extracted in S212: screen core features with high predictive power through correlation analysis and feature importance assessment (feature selection); construct new features such as ratios and differences by combining business knowledge (feature construction); perform one-hot encoding and label encoding conversion on categorical features (feature encoding); perform standardization / normalization processing on features (feature scaling) to eliminate the influence of different units; S222 Feature Library Structure Definition: Defines a structured classification standard for "basic features (static data such as enterprise registration information), behavioral features (dynamic data such as business activities), and environmental features (external data such as policy environment)". S223 Enterprise Feature Library Construction: The optimized features are organized into a feature matrix according to the above classification to complete the construction of a multi-dimensional enterprise feature library, realizing the structured storage and reuse of features.

[0042] Reference Figure 3 S23 Enterprise Classification: Building upon the structured enterprise feature library established in S22, its core is to achieve precise enterprise grouping through a progressive process of "clarifying classification criteria → selecting representative samples → constructing core feature representations → adaptive clustering partitioning," ultimately forming homogeneous subgroups with converging features. This provides the core basis for subsequent S3 grouping modeling (personalized model construction adapted to different subgroup features). The specific implementation consists of four sub-steps: S231 Determine the classification criteria: Select equity structure, life cycle stage, location, business scale, industry position, and industry from the enterprise feature database as the core classification criteria, comprehensively covering key dimensions such as enterprise static attributes, development stage, spatial distribution, and industry positioning; S232 Selecting a Representative Subset: Based on the above classification criteria, the data in the enterprise feature library is sampled and screened to retain enterprise samples with typical characteristics under each dimension, forming a representative enterprise subset to ensure the universality and relevance of subsequent classification results; S233 constructs core feature representation: Combining the selected representative subset, integrating the feature data corresponding to "industry location, region, and life cycle stage", constructing a "industry-space-life cycle" ternary embedding as the core feature representation to characterize the core differences of enterprises; S234 Adaptive Clustering Partitioning: Adaptive density peak clustering algorithm is used to calculate the enterprise data density and sample distance under the core feature representation, automatically identify the cluster center and dynamically adjust parameters such as cluster radius and density threshold, and divide the enterprises into multiple homogeneous subgroups with similar features.

[0043] Reference Figure 4 S3, Cluster Modeling: To adapt to the characteristic differences of different enterprise groups and fully leverage the advantages of multi-dimensional algorithms, it is necessary to build a dedicated ensemble prediction model for each subgroup, which is divided into three stages: S31 Model Adaptation Configuration: To ensure a high degree of matching between the model and the data types of the subgroups, a dedicated integrated model architecture needs to be configured for each homogeneous subgroup. Specifically, for each homogeneous subgroup, an independent project intelligence prediction model is built. This model is an integrated prediction model, and the corresponding sub-model combination is adapted according to the feature data type distribution of the homogeneous subgroups. For example, the subgroups with a high proportion of structured data focus on configuring the LightGBM or XGBoost model, while the subgroups with rich text data strengthen the BERT model configuration to ensure that the model architecture is highly consistent with the data features.

[0044] S32 Sub-model Training: To fully leverage the advantages of different algorithms in processing specific data types, specialized sub-models need to be trained for multiple feature types. Specifically, these sub-models include the LightGBM or XGBoost model for processing structured data, the BERT model for processing text data, the GraphSAGE model for processing graph data, and the Transformer model for processing time-series data. Each sub-model is trained based on its corresponding feature data. The LightGBM and XGBoost models excel at capturing non-linear relationships in structured data, the BERT model can deeply mine semantic information in text, the GraphSAGE model can effectively extract correlation features from graph networks, and the Transformer model excels at capturing long-term and short-term dependencies in time-series data, ensuring that the value of each type of feature is fully realized.

[0045] S33 Integration and Fusion Processing: To integrate the advantages of multiple sub-models and improve prediction accuracy and stability, it is necessary to integrate the outputs of each sub-model through a scientific fusion strategy. Specifically, the output results of each sub-model are integrated through Bayes Ensemble, and the prediction probabilities of different sub-models are weighted and fused. The weights are dynamically adjusted according to the performance of each sub-model on the validation set to obtain an integrated prediction model specific to this homogeneous subgroup. This enables collaborative modeling of multimodal data, ensuring that the prediction accuracy of project type, location, demand, etc., remains stable above 90%, and improving the model's predictive relevance and reliability for specific categories of enterprise project intelligence.

[0046] Reference Figure 5 S4. Model Validation and Dynamic Adjustment: This involves conducting dual validation to assess model performance and dynamically adjusting the model based on the validation results to ensure it meets the specific requirements of its intended use. This process is divided into three stages: S41 Dual Validation Execution: To comprehensively evaluate model performance in terms of both stability and discriminative power, dual validation within and across classes is required. Specifically, for each homogeneous subgroup, the ensemble prediction model is subjected to dual validation using "intra-class 5-fold cross-validation + cross-class k-1-fold cross-validation" (where k is the number of homogeneous subgroups). Intra-class 5-fold cross-validation evaluates the model's predictive stability among similar enterprises by dividing the subgroup data into 5 parts and using them alternately as the validation set. Cross-class k-1-fold cross-validation uses the data from the other k-1 subgroups as the validation set to detect the model's ability to distinguish between different types of enterprises, thus avoiding "overfitting" or "confusing predictions" problems.

[0047] S42 Performance Evaluation: To quantify the model's distinguishing ability, core indicator thresholds need to be set for scientific judgment. Specifically, core performance evaluation indicators such as the outlier AUC value of the model are calculated. A preset threshold of 0.75 is set. If the outlier AUC value is higher than this threshold, it is determined that the model has insufficient distinguishing ability between different categories, and there is a risk of misclassifying outlier companies as the target category, which fails to meet the prediction requirement of "specialized for specific categories".

[0048] S43 Dynamic Optimization and Adjustment: To address the issue of insufficient model discrimination, model optimization is achieved through category merging or feature filtering. Specifically, when the model's discrimination is insufficient, a dynamic adjustment operation is triggered: if it is due to overly fine category division, two homogeneous subgroups with feature similarity higher than 80% are merged to reduce category redundancy; if it is due to feature redundancy, the core features in the enterprise feature library are re-filtered, retaining features that contribute highly to category discrimination; the "validation-evaluation-adjustment" process is iterated repeatedly until all models meet the "dedicated to specific categories" performance requirements, ensuring the accuracy and reliability of model predictions.

[0049] Reference Figure 6 S5, Project Intelligence Prediction and Decision Support, utilizes a trained model to predict enterprise projects, generates a complete prediction report, and provides decision support. Results are automatically pushed out when the confidence level meets the criteria, supporting scenarios such as investment attraction and investment. It is specifically divided into three stages: S51 Target Enterprise Data Processing: Input new enterprise data and perform operations such as format standardization, missing value imputation, error data correction, and outlier detection and handling according to the S1 preprocessing standard to ensure that the quality of the input data is consistent with the model training data, laying the foundation for accurate prediction; S52 Project Intelligence Prediction Calculation: Match the pre-processed new enterprise data to the corresponding homogeneous subgroup, input the integrated prediction model exclusive to the subgroup, calculate and output the core prediction results in real time, including enterprise project type probability, TOP-N park / plot recommendation, land area demand, factory type demand, talent demand, policy dependence, and calculate the prediction confidence level and compare it with the preset threshold. S53 Report Generation, Push, and Decision Support: Based on the forecast results, standardized forecast reports are generated, and interpretable technologies such as SHAP and LIME are introduced to analyze core influencing factors. Visual content such as land plot heatmaps and demand radar charts are added to enhance report readability and credibility. If the forecast confidence level exceeds a preset threshold, the report is automatically pushed to relevant users. Simultaneously, targeted decision support is provided based on the forecast results, including recommending suitable investment projects, suggesting optimal park locations, and providing directions for investment promotion interventions, directly serving the practical needs of government investment promotion and industrial investment.

[0050] Reference Figure 7The present invention also provides a machine learning-based enterprise project intelligence prediction system, the system comprising: The data acquisition and preprocessing module is used to perform the operations of S1, which involves acquiring the enterprise's raw data through a multi-source data acquisition interface, and then performing preprocessing processes such as format standardization, missing value imputation, error data correction, and outlier detection and processing to output high-quality preprocessed data, ensuring the reliability of data support for subsequent stages.

[0051] The feature extraction and classification module is used to perform the S2 operation, that is, to extract multi-dimensional features from preprocessed data and perform optimization processing such as feature selection, construction, encoding, scaling, etc., to build enterprise feature vectors and structured enterprise feature library. At the same time, based on the "industry-space-life cycle" ternary embedding feature representation, the adaptive density peak clustering algorithm is used to divide enterprises into multiple homogeneous subgroups. The cluster modeling module is used to perform the S3 operation, which is to configure a combination of sub-models that are adapted to the data type of each homogeneous subgroup, train each sub-model separately, and then construct an independent ensemble prediction model through Bayesian fusion to achieve personalized modeling of "one model per class".

[0052] The validation and optimization module performs the operations of S4, namely, conducting dual validation of the ensemble prediction model: intra-class 5-fold cross-validation and cross-class k-1-fold cross-validation; calculating the AUC value of the different classes and comparing it with a preset threshold to evaluate the discrimination; if the discrimination is insufficient, triggering the merging of homogeneous subgroups (feature similarity ≥ 80%) or re-selection of core features; and introducing a PSI drift monitoring mechanism to track changes in data distribution in real time. When the PSI value exceeds a preset threshold, an incremental learning mechanism is triggered to perform incremental training and parameter updates on the ensemble prediction model based on new data, thereby achieving dynamic optimization of the model and ensuring the continuous effectiveness of the model under scenarios of changing data distribution.

[0053] The prediction output module is used to perform the operations of S5, namely, to perform preprocessing on the target enterprise data in the same way as the training data; to match the ensemble model with the corresponding homogeneous subgroup, output prediction results such as project type probability, park recommendation, and land use demand, and calculate the confidence level; and to generate a prediction report with interpretable technology (SHAP / LIME) and visualization content based on the results, which is automatically pushed when the confidence level reaches the target, while providing targeted decision support for scenarios such as investment promotion and investment.

[0054] The present invention also provides a computer device, including a processor, a memory, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, it implements the machine learning-based enterprise project intelligence prediction method described in S1 to S5 above. The processor's computing power ensures the efficient execution of each stage, such as data preprocessing, feature extraction, model training, verification and adjustment, and prediction output. The memory enables stable storage of information such as original enterprise data, preprocessed data, enterprise feature library, model parameters, and prediction results, providing reliable hardware support for the large-scale operation of the entire prediction method and ensuring that the system can quickly respond to the real-time prediction needs of scenarios such as government investment promotion and industrial investment.

[0055] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the machine learning-based enterprise project intelligence prediction method described in S1 to S5 above. This computer-readable storage medium enables convenient storage, dissemination, and deployment of the prediction method. Users can quickly apply the prediction technology by reading the computer program in the storage medium without having to repeatedly develop the core algorithm, effectively reducing the technical application threshold and facilitating the widespread implementation and promotion of this enterprise project intelligence prediction technology in different user groups and application scenarios.

[0056] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0057] The implementation principle of this invention is as follows: Based on the closed-loop logic of "data integration - feature characterization - precise clustering - personalized modeling - verification and optimization - dynamic adaptation", it integrates heterogeneous data from multiple sources such as industry and commerce, finance and taxation, patents, recruitment, land use, electricity consumption, policies, and supply chains. Data cleaning is completed through format standardization, missing value imputation, error correction, outlier handling, and redundant information removal, breaking down traditional data silos. Then, multimodal features of time series, spatial, textual, and graph structures are extracted from the preprocessed data. After screening, construction, encoding, scaling, and optimization, a structured enterprise feature library containing basic, behavioral, and environmental features is constructed. Combined with business dimensions such as industry, supply chain position, scale, region, life cycle, and equity structure, a "industry-space-life cycle" ternary embedded feature representation is constructed. Finally, an adaptive density peak clustering algorithm is used to divide enterprises into multiple homogeneous groups. For each subgroup, a dedicated ensemble prediction model is adapted, integrating the outputs of LightGBM / XGBoost (structured data), BERT (text data), GraphSAGE (graphical data), and Transformer (time series data) and using Bayesian fusion to improve prediction specificity. Through dual validation of "5-fold within a class + k-1 fold across classes," if the AUC of a different class is higher than 0.75, class merging or feature filtering is triggered to ensure the model is "specific to its class." Simultaneously, SHAP and LIME interpretability technologies are introduced to generate visual charts, coupled with PSI drift monitoring and incremental learning mechanisms to achieve dynamic model updates with zero downtime. This solves the problems of traditional methods such as single analysis dimensions, low prediction accuracy, static models, and uninterpretable results, ultimately achieving accurate prediction of enterprise project direction, location selection, and resource allocation needs.

[0058] The embodiments described herein are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based method for predicting enterprise project intelligence, characterized in that, Includes the following steps: Collect enterprise data and process the enterprise data to obtain preprocessed data; Based on the preprocessed data, multi-dimensional enterprise features are extracted to construct enterprise feature vectors, and enterprises are divided into multiple categories based on the enterprise feature vectors. For each of the defined categories, an independent project intelligence prediction model is constructed and trained. For each category of project intelligence prediction model, intra-category and cross-category validations are performed, and the category division or the selection of enterprise characteristics is dynamically adjusted based on the validation results. Input the target company's data into the corresponding project intelligence prediction model, and output the project intelligence prediction result.

2. The enterprise project intelligence prediction method based on machine learning according to claim 1, characterized in that, The process of collecting enterprise data and processing the data to obtain preprocessed data specifically includes: Collect raw enterprise data from multiple pre-defined heterogeneous data sources; The original data of the enterprise is preprocessed; The preprocessing includes format normalization, missing value handling, error data correction, outlier handling, and redundant information removal.

3. The enterprise project intelligence prediction method based on machine learning according to claim 1, characterized in that, Based on the preprocessed data, multi-dimensional enterprise features are extracted and an enterprise feature vector is constructed, specifically including: From the preprocessed data, time series features, spatial features, text features, and graph structure features are extracted; The extracted features are processed through feature selection, feature construction, feature encoding, and feature scaling to build enterprise feature vectors and form an enterprise feature library.

4. The enterprise project intelligence prediction method based on machine learning according to claim 3, characterized in that, The steps for classifying enterprises into multiple categories are as follows: Based on the enterprise feature library, and by integrating multiple preset business dimensions, an industry-space-lifecycle ternary embedding is constructed as an enterprise feature representation. An adaptive density peak clustering algorithm is used to automatically divide enterprises into multiple homogeneous subgroups based on the enterprise feature representation.

5. The enterprise project intelligence prediction method based on machine learning according to claim 1, characterized in that, For each of the aforementioned categories, an independent project intelligence prediction model is constructed and trained, specifically as follows: An ensemble prediction model is constructed for each category, wherein the ensemble prediction model is configured to fuse the outputs of multiple sub-models that process multiple data types.

6. The enterprise project intelligence prediction method based on machine learning according to claim 5, characterized in that, The integrated prediction model integrates the outputs of the following sub-models through Bayesian fusion: LightGBM or XGBoost models are used for processing structured data; BERT model for processing text data; The GraphSAGE model is used for processing spectral data. And the Transformer model for processing time-series data.

7. The enterprise project intelligence prediction method based on machine learning according to claim 1, characterized in that, The step of dynamically adjusting the category division or the selection of enterprise characteristics based on the verification results includes: If the performance evaluation index of the project intelligence prediction model in cross-class verification is higher than the preset threshold, it will trigger the reclassification of the category or the re-screening of the enterprise characteristics.

8. A machine learning-based enterprise project intelligence prediction system, characterized in that, The enterprise project intelligence prediction method based on machine learning, applied to any one of claims 1-7, specifically includes: The data acquisition and preprocessing module is used to collect relevant enterprise data from multiple heterogeneous data sources and preprocess the collected data to obtain preprocessed data. The feature extraction and classification module is used to extract multi-dimensional enterprise features from the preprocessed data to construct enterprise feature vectors, and to classify enterprises into multiple categories based on the enterprise feature vectors. The cluster modeling module is used to build and train independent project intelligence prediction models for each of the categories. The verification and optimization module is used to perform intra-class and cross-class verification on the project intelligence prediction model for each category, and to trigger adjustments to the category division in the feature extraction and classification module or to trigger adjustments to the feature selection in the feature extraction and classification module based on the results of the cross-class verification. The prediction output module is used to input the target company's data into the corresponding project intelligence prediction model and output the project intelligence prediction result.

9. A computer device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the machine learning-based enterprise project intelligence prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the machine learning-based enterprise project intelligence prediction method as described in any one of claims 1-7.

Images