Invocation target enterprise screening method fusing constraint conditions and multi-objective optimization

Abstract

The invention relates to a constraint condition and multi-objective optimization fused investment attraction target enterprise screening method. The method comprises the steps of obtaining target industry field information and generating an initial target enterprise pool; obtaining regional resource constraint information and enterprise multi-dimensional information, and inputting the information into a constraint discrimination model to obtain a candidate enterprise set; inputting enterprise multi-dimensional information associated with the candidate enterprise set into the tendency evaluation model to obtain an enterprise comprehensive score set; obtaining optimization target information, and constructing a multi-target optimization function; obtaining optimization condition information, matching a corresponding multi-objective optimization algorithm and adjusting parameters; solving the multi-objective optimization function, generating an enterprise screening strategy scheme set, and generating an enterprise screening decision report; in conclusion, by fusing the regional resource constraint condition and the multi-target optimization mechanism, the automatic processing and objective evaluation of the enterprise screening process are realized, and the method has the advantage of improving the efficiency and objectivity of screening the investment attracting target enterprises.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a target enterprise screening method fusing constraint conditions and multi-objective optimization. BACKGROUND

[0002] In the current enterprise investment work, the screening of target enterprises mainly relies on two types of technical paths.

[0003] The first type is the traditional manual research and judgment mode. The investment officers need to spend a lot of time consulting enterprise directories, financial reports, market information and other scattered information sources, and make subjective judgments and preliminary contacts combined with personal experience. This way highly depends on the professional quality and experience accumulation of personnel, and the screening process is easily affected by subjective factors, resulting in poor consistency of results. At the same time, when facing massive enterprise data, the efficiency is extremely low, it is difficult to quickly complete the preliminary screening, and it is also impossible to scientifically and quantitatively evaluate the potential relocation or expansion intentions of enterprises, such as unable to systematically analyze the influence of dynamic factors such as enterprise strategic adjustment and market environment change.

[0004] The second type is the digital platform that has appeared in recent years. Such platforms usually integrate static data such as enterprise business information and intellectual property rights, and can perform preliminary filtering by setting some basic conditions (such as registered capital range, industry classification, enterprise size threshold). However, these platforms have significant limitations in function, and most of them only stay at the level of information retrieval and linear sorting, the screening dimension is too single, and they fail to deeply mine the complex reasons behind the enterprise relocation decision, such as ignoring key factors such as environment, resource matching degree, and industry chain synergy, and lack the ability to systematically match and optimize the resource constraints of the investment area with the needs of enterprises. SUMMARY

[0005] In order to solve the above-mentioned defects, the present application provides a target enterprise screening method for investment fusing constraint conditions and multi-objective optimization.

[0006] The above-mentioned invention purpose of the present application is realized by the following technical scheme: A target enterprise screening method for investment fusing constraint conditions and multi-objective optimization, comprising the steps of: obtaining target industry field information, screening a pre-constructed enterprise database based on the target industry field information, and generating an initial target enterprise pool according to the screening result; obtaining regional resource constraint information and enterprise multi-dimensional information associated with the initial target enterprise pool, and inputting the enterprise multi-dimensional information and the regional resource constraint information into a pre-trained constraint discrimination model to obtain a candidate enterprise set; inputting the enterprise multi-dimensional information associated with the candidate enterprise set into a pre-trained inclination evaluation model to obtain an enterprise comprehensive score set; obtaining optimization target information, and constructing a multi-objective optimization function based on the enterprise comprehensive score set and the optimization target information; obtaining optimization condition information, matching a corresponding multi-objective optimization algorithm from a pre-constructed algorithm configuration library based on the optimization condition information and the multi-objective optimization function, and adjusting parameters of the multi-objective optimization algorithm; solving the multi-objective optimization function by the adjusted multi-objective optimization algorithm, generating an enterprise screening strategy scheme set, and generating an enterprise screening decision report based on the enterprise screening strategy scheme set.

[0007] In a preferred example, the application can be further configured to: the step of obtaining target industry field information, filtering the pre-constructed enterprise database based on the target industry field information, and generating an initial target enterprise pool according to the filtering result, comprising the steps of: performing natural language processing and semantic extension on the target industry field information to generate an industry feature vector; extracting basic information data of each enterprise from the enterprise database, and constructing an enterprise portrait feature set of each enterprise based on a pre-defined feature engineering rule library; extracting enterprise meta-features based on the enterprise portrait feature set, and calculating initial weights of the enterprise portrait feature set based on the enterprise meta-features; performing similarity calculation on the enterprise portrait feature set and the industry feature vector, sorting each enterprise according to the similarity calculation result, and generating the initial target enterprise pool.

[0008] In a preferred example, the application can be further configured to: the constraint discrimination model comprises a graph construction layer, a graph convolution layer and a readout layer, the step of obtaining regional resource constraint information and enterprise multi-dimensional information associated with the initial target enterprise pool, and inputting the enterprise multi-dimensional information and the regional resource constraint information into the pre-trained constraint discrimination model to obtain a candidate enterprise set, comprising the steps of: the graph construction layer constructs an enterprise association information graph based on the enterprise multi-dimensional information, wherein the nodes of the enterprise association information graph are enterprises, and the association relationship is an edge; the graph convolution layer updates the node features of each node in the enterprise association information graph by performing information transmission and node feature update on the enterprise association information graph, so that the node features of each node in the enterprise association information graph include multi-hop neighborhood structure information of the enterprise corresponding to the node; the readout layer fuses the regional resource constraint information and the node features in the enterprise association information graph, performs global pooling and node-level discrimination on the fusion features generated after fusion, and generates the candidate enterprise set based on the node-level discrimination result.

[0009] The application can be further configured in a preferred example that: the graph convolution layer makes the node feature of each node in the enterprise association information graph include the multi-hop neighborhood structure information of the enterprise corresponding to the node by information transmission and node feature updating on the enterprise association information graph, and the step includes the steps of: (a) generating the corresponding message vector for each edge in the enterprise association information graph; (b) aggregating the message vectors corresponding to the edges pointing to each node in the enterprise association information graph to generate the aggregated adjacent message vector of each node; (c) fusing the aggregated adjacent message vector of each node in the enterprise association information graph with the current feature representation of the node to generate the updated feature representation of the node and replace the current feature representation of the node, wherein the current feature representation of the node is the enterprise multi-dimensional information when the fusion is performed for the first time; (d) taking steps (a), (b), and (c) as an iteration, repeating the iteration until a termination condition is reached, so that the node feature of each node in the enterprise association information graph includes the multi-hop neighborhood structure information of the enterprise corresponding to the node, and the termination condition includes that the number of iterations reaches a preset iteration number K.

[0010] The application can be further configured in a preferred example that: the inclination evaluation model includes a feature encoding layer, a multi-task learning layer, and a fusion output layer, and the step of inputting the enterprise multi-dimensional information associated with the candidate enterprise set into the pre-trained inclination evaluation model to obtain the enterprise comprehensive score set includes the steps of: The feature encoding layer extracts and encodes the enterprise multi-dimensional information to generate the deep feature representation corresponding to each enterprise in the candidate enterprise set; The multi-task learning layer performs collaborative learning on the deep feature representation through a parallel main task network and a plurality of auxiliary task networks to generate the main task score and the auxiliary task score corresponding to each enterprise in the candidate enterprise set; The fusion output layer performs weighted fusion on the main task score and the auxiliary task score corresponding to each enterprise to generate the enterprise comprehensive score set.

[0011] The application can be further configured in a preferred example that: the multi-task learning layer performs collaborative learning on the deep feature representation through a parallel main task network and a plurality of auxiliary task networks to generate the main task score and the auxiliary task score corresponding to each enterprise in the candidate enterprise set, and the step includes the steps of: inputting the deep feature representation into the main task network and each auxiliary task network to generate the initial main score and a plurality of initial auxiliary scores through single forward propagation; The score consistency optimization function is constructed with the initial main scores and the initial auxiliary scores as inputs, and the optimization objective is to minimize the difference between the weighted sum of the initial main scores and the weighted sum of the initial auxiliary scores, while minimizing the variance between the initial auxiliary scores; The main task scores and the auxiliary task scores are generated by solving the score consistency optimization function.

[0012] In a preferred example, the application can be further configured to: the step of obtaining the optimization target information and constructing the multi-objective optimization function based on the enterprise comprehensive score set and the optimization target information comprises the steps of: The optimization target information is parsed to extract N optimization sub-targets and the priority weights corresponding to each optimization sub-target; The value distribution statistics of each optimization sub-target corresponding to the objective function on the candidate enterprise set are calculated based on the enterprise comprehensive score set, and the normalized objective function is generated by normalizing each objective function according to the value distribution statistics. The regional resource constraint information is parsed to obtain rigid constraints and soft constraints; The rigid constraints are converted into a hard constraint inequality set, the corresponding penalty function is constructed based on the soft constraints, and the constructed penalty function is taken as the N+1 optimization sub-target; The normalized objective function and the penalty function are combined based on the priority weights of each optimization sub-target, and the multi-objective optimization function is constructed based on the combination result and the hard constraint inequality set.

[0013] In a preferred example, the application can be further configured to: the step of obtaining the optimization condition information, matching the corresponding multi-objective optimization algorithm from the pre-constructed algorithm configuration library based on the optimization condition information and the multi-objective optimization function, and adjusting the parameters of the multi-objective optimization algorithm comprises the steps of: The optimization feature vector is extracted based on the multi-objective optimization function and the hard constraint inequality set, and the optimization feature vector includes the optimization problem size, the constraint tightness, the objective function nonlinearity, and the optimization target correlation; The optimization feature vector is matched with the historical cases in the pre-constructed algorithm configuration library, and the historical cases each include a historical optimization feature vector and a corresponding preferred algorithm type and parameter configuration; The target historical case is determined according to the similarity matching result, and the preferred algorithm type corresponding to the target historical case is taken as the basic algorithm framework; The feature difference vector of the optimization feature vector and the historical optimization feature vector of the target historical case is calculated, and the parameter configuration of the basic algorithm framework is adjusted based on the preset parameter adjustment strategy and the feature difference vector.

[0014] The second application object of the present application is achieved by the following technical solutions. A target enterprise screening system fusing constraint conditions and multi-objective optimization, comprising: A screening module configured to obtain target industry field information, screen a pre-constructed enterprise database based on the target industry field information, and generate an initial target enterprise pool according to a screening result; A first input module configured to obtain regional resource constraint information and enterprise multi-dimensional information associated with the initial target enterprise pool, and input the enterprise multi-dimensional information and the regional resource constraint information into a pre-trained constraint discrimination model to obtain a candidate enterprise set; A second input module configured to input enterprise multi-dimensional information associated with the candidate enterprise set into a pre-trained inclination evaluation model to obtain an enterprise comprehensive score set; An optimization function construction module configured to obtain optimization target information, and construct a multi-objective optimization function based on the enterprise comprehensive score set and the optimization target information; An optimization algorithm adjustment module configured to obtain optimization condition information, match a corresponding multi-objective optimization algorithm from a pre-constructed algorithm configuration library based on the optimization condition information and the multi-objective optimization function, and adjust parameters of the multi-objective optimization algorithm; A decision generation module configured to solve the multi-objective optimization function by using the adjusted multi-objective optimization algorithm, generate a set of enterprise screening strategy schemes, and generate an enterprise screening decision report based on the set of enterprise screening strategy schemes.

[0015] In a preferred example, the present application can be further configured as follows: A feature vector generation submodule configured to perform natural language processing and semantic extension on the target industry field information to generate an industry feature vector; A portrait feature construction submodule configured to extract basic information data of each enterprise from the enterprise database, and construct an enterprise portrait feature set of each enterprise based on a pre-defined feature engineering rule library; An initial weight calculation submodule configured to extract enterprise meta-features based on the enterprise portrait feature set, and calculate initial weights of the enterprise portrait feature set based on the enterprise meta-features; A similarity calculation submodule configured to perform similarity calculation on the enterprise portrait feature set and the industry feature vector, sort each enterprise according to a similarity calculation result, and generate the initial target enterprise pool.

[0016] In summary, the application provides a target enterprise screening method for investment, which fuses constraint conditions and multi-objective optimization. The method realizes automatic processing and objective evaluation of the enterprise screening process, effectively integrates dynamic matching of enterprise multi-dimensional information and regional resource endowment, supports multi-dimensional optimization analysis to generate a scientific decision report, and has the advantages of improving the efficiency and objectivity of target enterprise screening for investment, realizing systematic dynamic matching of regional resource constraints and enterprise needs, and supporting multi-dimensional optimization analysis to generate a scientific decision. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of an embodiment of the target enterprise screening method for investment fusing constraint conditions and multi-objective optimization of the application; Figure 2 is an implementation flowchart of step S10 in the embodiment of the target enterprise screening method for investment fusing constraint conditions and multi-objective optimization of the application; Figure 3 is an implementation flowchart of step S20 in the embodiment of the target enterprise screening method for investment fusing constraint conditions and multi-objective optimization of the application. DETAILED DESCRIPTION

[0018] The following will be described in detail in combination with the accompanying drawings. Figures 1-3 The application will be further described in detail.

[0019] In an embodiment, as shown in the accompanying drawings, Figure 1 The application discloses a target enterprise screening method for investment fusing constraint conditions and multi-objective optimization, which specifically includes the following steps: S10: Obtain target industry field information, screen a pre-constructed enterprise database based on the target industry field information, and generate an initial target enterprise pool according to the screening result; In this embodiment, the target industry field information refers to the industrial development direction set by the investment subject, which is the primary condition for enterprise screening. Its form can be a standard industry classification code, a keyword list (such as "new energy vehicles", "biological medicine", etc.), or a specific industry chain link name (such as "chip design", "high-end packaging", etc.). The target industry field information is used to quickly locate the relevant industry category from a large number of enterprises and is the basis for subsequent fine screening. The pre-constructed enterprise database refers to a structured data set that stores enterprise-related information. It can be implemented as a relational database system, such as using MySQL to store enterprise business registration information and financial data, or using a NoSQL database such as MongoDB to store unstructured enterprise-related information, thereby providing a data basis for initial screening. The initial target enterprise pool refers to the enterprise set obtained by screening the pre-constructed enterprise database based on the target industry field information.

[0020] S20: Obtain regional resource constraint information and initial target enterprise pool associated enterprise multi-dimensional information, and input the enterprise multi-dimensional information and the regional resource constraint information into the pre-trained constraint discrimination model to obtain a candidate enterprise set; In the embodiment, the regional resource constraint information refers to various restrictive conditions and boundaries followed by the investment subject to achieve the investment target in the investment decision-making. The regional resource constraint information includes resource bearing constraints and strategy-oriented constraints. The resource bearing constraints refer to the limited physical resources and rules and regulations resources objectively existing in the local, such as available land, subsidy budget, environmental capacity, energy indicators, and talent upper limit. The strategy-oriented constraints refer to the directional and rule-based restrictions set based on the regional development strategy, such as the prohibited industry catalog (such as high-pollution and high-energy consumption industries), preference for enterprise type, and investment intensity requirement (such as the investment amount per mu of land should not be lower than a certain standard). The enterprise multi-dimensional information refers to a multi-source and multi-type data set for describing the state of an enterprise in multiple aspects, which goes beyond the traditional business information (such as registered capital and establishment time) and is a dynamic and deep enterprise portrait data. The pre-trained constraint discrimination model refers to a discriminant model for processing regional resource constraints and enterprise multi-dimensional information, which can be implemented by using a support vector machine classifier, such as processing nonlinear constraint boundaries by using a radial basis kernel function or performing constraint condition discrimination based on feature importance by using a random forest model. Specifically, the constraint discrimination model generates a candidate enterprise set by inputting the enterprise multi-dimensional information and the regional resource constraint information. The candidate enterprise set refers to the enterprise set meeting the constraint conditions output by the constraint discrimination model after filtering the initial target enterprise pool.

[0021] Further, the enterprise multi-dimensional information usually includes: Business dimension: financial data (revenue, profit), personnel size, recruitment trend, and intellectual property output; Development dimension: outward investment record, capacity utilization rate, new project announcement, and supply chain location; Dynamic information flow dimension: expansion, relocation, cooperation, and other dynamic information mentioned in news reports and industry research reports; Risk dimension: compliance and credit record information, business stability representation, negative behavior event record, and disposal record information faced by the enterprise in business.

[0022] S30: Input the enterprise multi-dimensional information associated with the candidate enterprise set into the pre-trained inclination evaluation model to obtain an enterprise comprehensive score set; In the embodiment, the pre-trained propensity evaluation model refers to an evaluation model for quantifying the enterprise comprehensive score or the intensity of spillover transfer propensity. The propensity evaluation model analyzes the multi-dimensional information of an enterprise, comprehensively calculates the performance of the enterprise in multiple dimensions (such as enterprise strength, spillover driving force, and transfer ability), and fuses them into a comparable comprehensive index. The enterprise comprehensive score set refers to a set of standardized and comparable evaluation index results obtained by quantitatively evaluating each enterprise in the candidate enterprise set by using the propensity evaluation model.

[0023] S40: Obtain optimization target information, and construct a multi-objective optimization function based on the enterprise comprehensive score set and the optimization target information. In the embodiment, the optimization target information refers to a quantitative description of the core intention and strategic orientation of the investment subject in the investment activity. Generally, the optimization target information is not a single target, but a set of usually interrelated and even possibly conflicting targets, which are used to guide the multi-objective optimization algorithm to find the optimal enterprise combination scheme. The multi-objective optimization function refers to a mathematical expression of the investment target, which constructs multiple conflicting investment targets into independent mathematical functions (i.e., objective functions), and these mathematical functions constitute the target vector that needs to be optimized.

[0024] S50: Obtain optimization condition information, match the corresponding multi-objective optimization algorithm from the pre-constructed algorithm configuration library based on the optimization condition information and the multi-objective optimization function, and adjust the parameters of the multi-objective optimization algorithm. In the embodiment, the optimization condition information refers to external environmental factors and performance requirements that need to be considered in addition to the problem itself defined by the multi-objective optimization function when starting optimization. The pre-constructed algorithm configuration library refers to a reference set for storing optimization algorithm parameter configurations, which can be implemented by using a rule-based matching system, such as mapping optimization problem characteristics to fixed algorithm parameters through pre-defined conditional statements, or using a simple lookup mechanism of historical cases to select standard algorithm configurations according to problem size, thereby supporting parameter adjustment of the multi-objective optimization algorithm.

[0025] S60: Solve the multi-objective optimization function by using the adjusted multi-objective optimization algorithm, generate a set of enterprise screening strategy schemes, and generate an enterprise screening decision report based on the set of enterprise screening strategy schemes.

[0026] In this embodiment, the adjusted multi-objective optimization algorithm refers to an instantiated optimization algorithm instance that has been type-selected and parameter-optimized for the current specific optimization problem; the enterprise screening strategy scheme set refers to the direct output of the multi-objective optimization algorithm, which is usually a Pareto optimal solution set, i.e., not a single optimal scheme, but a set of non-dominated solutions that have different trade-offs between multiple objectives. For example, scheme A focuses on introducing high-quality enterprises (highest quality score, but medium total investment); scheme B focuses on introducing large investment projects (largest total investment, but average quality score is medium); scheme C is a balanced scheme in terms of quality and investment; and the enterprise screening decision report refers to the final output document for investment decision makers generated after post-processing, analysis and visualization of the enterprise screening strategy scheme set.

[0027] Further, the enterprise screening decision report usually includes: Scheme summary: clearly list the core indicators of several main alternative schemes, such as total score, total investment, and employment promotion, etc. Scheme details: specific enterprise list, recommendation reasons, and resource demand analysis for each scheme; Risk and benefit assessment: explanation of potential risks and expected benefits for each scheme; Visualized charts: such as Pareto frontier chart, which intuitively shows the trade-off relationship between schemes.

[0028] Specifically, the present application provides a systematic method for intelligent screening of target enterprises for investment. First, based on the set target industry, an initial target enterprise pool is automatically generated from the enterprise database, completing efficient preliminary screening. Then, by introducing regional resource constraints, the pre-trained constraint discrimination model is used to intelligently identify whether the enterprises meet the local development constraints, and a candidate enterprise set is screened out to ensure the feasibility of subsequent analysis targets in terms of resources. On this basis, the pre-trained inclination evaluation model is further used to conduct multi-dimensional potential analysis of the candidate enterprises, generating a set of quantitative enterprise comprehensive scores, thereby converting complex characteristics such as enterprise spillover inclination into objective and comparable indicators. Next, based on the multiple objectives of the investment party (such as improving quality, expanding investment, and enhancing industrial synergy) and the enterprise comprehensive score set, a multi-objective optimization function is constructed, and the optimal multi-objective optimization algorithm is intelligently matched from the algorithm library according to the characteristics of the current problem and is parameter-optimized. Finally, through optimization solving, a set of investment strategy schemes that achieve the best balance between multiple objectives is generated, and a decision report with detailed analysis and comparison is automatically generated, providing clear and scientific decision-making basis for investment personnel.

[0029] Exemplarily, as a specific implementation, when the investment area focuses on the new energy vehicle industry, the target industry field information is set as "new energy vehicle", the enterprises related to battery manufacturing, motor research and development, etc. are screened out from the enterprise database to generate an initial target enterprise pool; the regional resource constraint information such as land supply index and environmental protection regulation requirements is input into the constraint discrimination model, and the constraint discrimination model outputs a candidate enterprise set that meets the regional carrying capacity by combining the enterprise multi-dimensional information (including capacity data, supply chain relationship, etc.); the enterprise multi-dimensional information of the candidate enterprise set is input into the inclination evaluation model to generate an enterprise comprehensive score including investment intention and technology matching degree; a multi-objective optimization function is constructed based on optimization target information such as investment amount priority and employment contribution weight, etc., an improved NSGA-II algorithm is matched based on optimization condition information through an algorithm configuration library and the population size parameter is adjusted, and a set of enterprise screening strategy schemes that take into account economic benefits and sustainable development are generated after solving.

[0030] Through the above technical solutions, the scheme effectively solves the problems of low efficiency of traditional manual screening and single screening dimension of existing digital platforms, realizes systematic consideration of complex reasons for enterprise relocation decision through dynamic matching of enterprise associated network by the constraint discrimination model; at the same time, the construction of multi-objective optimization function and the algorithm adaptive adjustment mechanism convert the regional resource constraints and enterprise demand into quantifiable analysis optimization problems, significantly improve the accuracy and decision support ability of the screening process, and avoid the limitations of manual experience dependence and single dimension screening.

[0031] In an embodiment, as shown in Figure 2 Step S10 includes: S11: natural language processing and semantic extension are performed on the target industry field information to generate an industry feature vector; In this embodiment, natural language processing refers to a technology for computers to understand, interpret and operate human language, in this step, natural language processing is used to parse the target industry field information (such as "new energy vehicle battery management system") and decompose it into core keywords and semantic units; semantic extension refers to a technology for automatically associating and adding synonyms, near synonyms and upstream and downstream industry terms based on semantic understanding, such as extending from "battery management system" to "BMS", "battery cell", "battery pack" and "thermal management", etc., which ensures the multi-dimensionality of the screening range and reduces the omission of related enterprises due to different terminology expressions; the industry feature vector refers to a technology for converting industry description into a high-dimensional numerical vector through text semantic analysis, which can be realized by using a pre-trained language model such as BERT or Word2Vec, and the purpose is to capture the implicit semantic association and dynamic evolution characteristics of the industry field, avoiding the isolation of keyword matching.

[0032] S12: Extract the basic information data of each enterprise from the enterprise database, and construct the enterprise portrait feature set of each enterprise based on the pre-defined feature engineering rule library; In this embodiment, the basic information data of the enterprise refers to a standardized and structured data set for uniquely identifying the enterprise entity and describing its fundamental attributes, which generally includes: identification information (unified social credit code, enterprise name, registration number, etc.), basic attribute information (registration address, main business scope, industry to which it belongs, enterprise type, date of establishment, registered capital, etc.), state information (operation status, registered capital, etc.), structure information (shareholder information, main management personnel, etc.); the pre-defined feature engineering rule library refers to a rule set encapsulating data conversion logic and domain knowledge, which defines how to standardize and normalize the original and scattered enterprise basic information data, and construct features with statistical significance and predictive ability, for example, the rules can include: "convert'registered capital' to logarithmic scale to eliminate dimensional influence", "perform word segmentation and TF-IDF vectorization on 'business scope' text", "divide 'establishment period' into'start-up period', 'growth period','mature period', etc. categories"; the enterprise portrait feature set can be understood as a structured data set integrating the multi-dimensional attributes of the enterprise, which can be based on heterogeneous data sources such as business registration, operation dynamics and intellectual property rights, and feature extraction and conversion can be achieved through the feature engineering rule library, and the purpose is to systematically construct a complete representation reflecting the comprehensive attributes of the enterprise, wherein the portrait feature set of an enterprise may include its size characteristics, technical characteristics, operating stability characteristics, etc.

[0033] S13: Extract enterprise meta-features based on the enterprise portrait feature set, and calculate the initial weights of the enterprise portrait feature set based on the enterprise meta-features; In this embodiment, the enterprise meta-feature refers to a high-level abstract feature that describes the inherent statistical characteristics of the enterprise portrait feature set, which can be realized by using statistical quantities such as mean, variance or information entropy of feature distribution, and the purpose is to quantitatively evaluate the discrimination ability and stability of each feature dimension; the initial weight refers to the feature importance coefficient dynamically assigned according to the enterprise meta-feature, which can be realized by using the entropy weight method or the regression model based on historical data, and the purpose is to adaptively adjust the weight according to the actual discriminant value of the feature, to solve the screening bias caused by fixed weight, for example, for an enterprise with high data completeness and strong feature discrimination, its portrait feature set can be assigned a higher initial weight when calculating, because its data is more reliable and the information quantity is larger; on the contrary, for an enterprise with serious data missing, its initial weight will be reduced.

[0034] S14: Calculate the similarity between the enterprise portrait feature set and the industry feature vector, sort each enterprise according to the similarity calculation result, and generate an initial target enterprise pool.

[0035] In the embodiment, the similarity calculation refers to a mathematical operation for quantifying the matching degree of the enterprise portrait feature set and the industry feature vector, which can be realized by using cosine similarity or Mahalanobis distance, and the purpose is to accurately evaluate the matching degree of the enterprise and the target industry.

[0036] Specifically, the scheme of the present application realizes deep semantic analysis of industry domain information through the generation of the industry feature vector, so that the text description is converted into a computable vector space representation. On this basis, the construction of the enterprise portrait feature set maps the enterprise static data into a structured feature vector, ensuring the systematic integration of multi-source information. Further, the extraction of enterprise meta-features and the calculation of initial weights form a dynamic weight adjustment mechanism, which automatically optimizes the contribution proportion of each dimension according to the statistical characteristics of the features. Finally, the enterprise portrait feature set and the industry feature vector are matched through similarity calculation, and the initial target enterprise pool is generated according to the quantitative results, thereby realizing a complete screening process from semantic expansion to dynamic weighting. Through this process, the generation process of the initial enterprise pool has semantic adaptability and weight self-adaptability.

[0037] Exemplarily, as a specific implementation, in generating the industry feature vector, a pre-trained BERT model is used to perform word segmentation and embedding processing on the industry description text, and an extended semantic vector is generated through synonym library expansion. In constructing the enterprise portrait feature set, basic information such as registered capital, industry code, number of patents in the past three years, and employee size is extracted from the enterprise database, and discrete fields are converted into numerical feature vectors according to the feature engineering rule library. In calculating the initial weight, the variance of each feature dimension is extracted as a meta-feature based on the enterprise portrait feature set, and the entropy weight method is used to calculate the weight of the feature with larger variance. In the similarity calculation stage, the cosine similarity is used to measure the angle between the enterprise feature vector and the industry feature vector, and the initial target enterprise pool is generated by sorting the top 500 enterprises according to the similarity score from high to low.

[0038] Through the above technical scheme, the present application effectively solves the problems of rough information representation and weight imbalance in the initial enterprise screening, significantly reduces the noise data in the initial target enterprise pool, and completely retains the high-potential target enterprises, thereby providing a high-quality input basis for subsequent constraint discrimination and multi-objective optimization.

[0039] In an embodiment, as shown in Figure 3 The constraint discrimination model includes a graph construction layer, a graph convolution layer, and a readout layer, and step S20 includes: S21: The graph construction layer constructs an enterprise correlation information graph based on enterprise multi-dimensional information, wherein the nodes of the enterprise correlation information graph are enterprises, and the correlation relationship is an edge. In this embodiment, the graph construction layer refers to a component layer for converting enterprise multi-dimensional information into a structured graph representation, which can be constructed based on supply chain data, industry similarity, or geographical proximity, etc. The purpose is to explicitly express the potential connection between enterprises. The enterprise correlation information graph refers to a data structure for representing the correlation between enterprises. In the enterprise correlation information graph, each node represents an independent enterprise entity, and the edge represents the existence of a specific type of relationship between enterprises. The purpose is to capture the network topology in the business ecosystem.

[0040] S22: The graph convolution layer updates the node features of each node in the enterprise correlation information graph by performing information transmission and node feature updating on the enterprise correlation information graph, so that the node features of each node in the enterprise correlation information graph include multi-hop neighborhood structure information of the enterprise corresponding to the node. In this embodiment, the graph convolution layer specifically refers to a component layer that performs graph neural network operations. It can use message passing mechanisms or spectral domain convolution to realize information propagation on the graph structure. The purpose is to make the node features contain multi-hop neighborhood information. Information transmission is a core computing mechanism in graph neural networks that simulates the way humans transmit social information. Its essence is that each node (enterprise) in the graph receives information from its neighbor nodes through the edges (relationships) directly connected to it, integrates and updates the information with its own information, and forms a new representation that is more comprehensive and richer. Node features refer to numerical vectors attached to each node (enterprise) in the graph. The numerical vectors come from the results of preliminary encoding of enterprise multi-dimensional information of the enterprise, and represent the mathematical representation of the enterprise in a high-dimensional feature space. Multi-hop neighborhood structure information refers to the information that each node's features not only contain its own attributes, but also aggregate information from its neighbors (one-hop), neighbors' neighbors (two-hop), and even more distant (multi-hop) nodes, representing the context information of the node in its local network environment. For example, by capturing multi-hop neighborhood information, potential supply chain crises can be warned in the case where a key supplier (two-hop neighbor) of an enterprise exists in high risk, even though the enterprise itself is of high quality, achieving deep insights beyond individual enterprise analysis.

[0041] S23: The readout layer fuses the regional resource constraint information with the node features in the enterprise correlation information graph, globally pools the fusion features generated after fusion, and performs node-level discrimination based on the node-level discrimination results to generate a candidate enterprise set.

[0042] In the embodiment, the readout layer refers to a component layer for integrating graph-level information and generating a final output. An attention mechanism or a pooling operation can be used to fuse the regional resource constraint information and the node features, with the purpose of realizing the coupling discrimination of the constraint conditions and the enterprise network features. The fused features refer to new feature representations generated after combining the regional resource constraint information and the node features. Instead of judging the enterprises in isolation, the discrimination is performed in a specific local resource constraint background. For example, a high-quality enterprise with a large land demand may not be suitable in a region with scarce land resources, but it can be the best target in a region with abundant resources. The fused features enable the discrimination process to be based on situational awareness. Global pooling refers to an operation of aggregating all node features in the graph, such as taking the average or maximum value of all node features. The purpose is to obtain a global representation of the entire graph structure, which is used for graph-level tasks, such as judging the overall risk of the entire enterprise group. Node-level discrimination is opposite to global pooling, which refers to classifying or regressing each node (enterprise) in the graph individually to determine whether it meets the conditions, which is a direct basis for generating a candidate enterprise set.

[0043] Specifically, the present scheme converts enterprise multi-dimensional information into a structured graph representation through the graph construction layer, uses the graph convolution layer to iteratively propagate node features on the graph structure to capture multi-hop influence effects, and finally fuses the regional resource constraint information and the graph features through the readout layer for global evaluation and node screening, thereby realizing accurate screening of candidate enterprises based on consideration of the network dependency between enterprises.

[0044] Illustratively, as a preferred embodiment, the graph construction layer constructs an enterprise association information graph based on the supply chain data in the enterprise business information, where the nodes are enterprises and the edges represent the supplier-customer relationship. The graph convolution layer uses a graph attention network for feature propagation, so that the features of each enterprise node contain three layers of neighborhood information. The readout layer concatenates the regional land resource constraint information and the node features, and then performs node-level discrimination through a fully connected network to output a candidate enterprise set.

[0045] Through the above technical solutions, the present application effectively solves the problem that the network dependency is ignored in the enterprise relocation decision, so that the candidate enterprise screening result can systematically consider indirect correlation factors such as supply chain collaboration and industrial cluster effect, thereby improving the accuracy and adaptability of the investment strategy.

[0046] In an embodiment, step S22 includes: (a) generating a message vector corresponding to each edge in the enterprise association information graph; (b) aggregating the message vectors corresponding to the edges pointing to each node in the enterprise association information graph to generate an aggregated adjacent message vector for each node; (c) fusing the aggregated adjacent message vectors of all nodes in the enterprise associated information graph with their current feature representations to generate updated feature representations of the nodes and replace their current feature representations, wherein the current feature representation of a node is the enterprise multidimensional information when the fusion is performed for the first time; (d) taking steps (a), (b), (c) as an iteration, repeating the iteration until a termination condition is reached, so that the node features of each node in the enterprise associated information graph include the multi-hop neighborhood structure information of the enterprise corresponding to the node, and the termination condition includes that the number of iterations reaches a preset iteration number K.

[0047] In the embodiment, the message vector refers to a feature carrier data unit that carries the association strength in the information transmission process, can be implemented by vector mapping based on edge attributes or neural network transformation, and aims to customize the expression of local network structure features according to the actual business association between enterprises; the aggregated adjacent message vector refers to a set of aggregated message vectors transmitted by all neighbor nodes, can be implemented by weighted average, attention mechanism or maximum pooling, and aims to integrate the contributions of multi-directional neighborhood information and avoid one-sidedness of single neighborhood information; the updated feature representation refers to an updated feature representation generated by each node after obtaining the aggregated adjacent message vector, in combination with the current feature representation thereof, and aims to ensure that the feature update gradually incorporates context information from the original business data and strengthens semantic coherence; the iteration refers to a cyclic mechanism of repeated information transmission and update, can be implemented by fixed number of iterations or dynamic termination based on a convergence threshold, and aims to capture high-order association relationships through multiple information propagation, while balancing the calculation efficiency and depth, and each iteration propagates the information by one hop (one layer) in the graph, and the iteration number K is a key hyperparameter that directly controls the range of information propagation, i.e., the size of the receptive field; when K = 1, the node only contains the information of its direct neighbor (one hop); when K = 2, the node contains the information of two-hop neighbors; through preset K iterations, it can be ensured that the final feature of each node contains the structure information in the K-hop neighborhood of the node, thereby realizing deep perception of complex association relationships.

[0048] Specifically, the scheme first generates a message vector for each edge in the enterprise correlation information graph, so that the information transmission is customized based on the industry similarity and business correlation between enterprises; then, the node aggregates the message vectors corresponding to the edges pointing to the node, integrates the contributions of multi-directional neighbor information to form an aggregated adjacent message vector; then, the aggregated adjacent message vector and the current feature representation of the node are fused to generate an updated feature representation and replace the original feature, ensuring that the first iteration uses enterprise multi-dimensional information as the initial input; finally, steps (a), (b) and (c) are repeatedly executed until the preset iteration number K is reached, so that the node feature gradually absorbs the multi-hop neighborhood structure information; the above sequential execution and iteration mechanism realize the systematic propagation of information from local to global, ensuring that the node features in the enterprise correlation information graph reliably integrate multi-hop neighborhood information, thereby effectively supporting the accurate decision-making of the candidate enterprise screening by the constraint discrimination model.

[0049] Exemplarily, as a preferred embodiment, in the target enterprise screening process, after the enterprise correlation information graph is constructed, the graph convolution layer first calculates a message vector for each edge, for example, generates a vector based on the enterprise industry classification code similarity and the supply chain correlation; then, for each node, the message vectors of all input edges are aggregated, and a weighted average method based on the enterprise size weight is used to generate an aggregated adjacent message vector; then, the aggregated adjacent message vector and the current feature of the node (initially the enterprise multi-dimensional information) are fused through a gated recurrent unit to generate an updated feature representation; finally, the above process is repeated K times, where K is set to 3, so that the node feature contains three-hop neighborhood information, which is used for subsequent candidate enterprise set generation.

[0050] Through the above scheme, the application can systematically capture the complex multi-hop correlation between enterprises, improve the accuracy of candidate enterprise screening, and improve the robustness of the constraint discrimination model.

[0051] In an embodiment, the tendency evaluation model includes a feature encoding layer, a multi-task learning layer, and a fusion output layer, and step S30 includes: S31: The feature encoding layer performs feature extraction and feature encoding on the enterprise multi-dimensional information to generate a deep feature representation corresponding to each enterprise in the candidate enterprise set; In this embodiment, the feature encoding layer refers to a component layer that performs original data to high-order semantic feature conversion, which can be implemented using a convolutional neural network or a recurrent neural network architecture. Specifically, it can extract the time sequence association and structured pattern in the multi-dimensional information of an enterprise through an embedding layer and a multi-layer nonlinear transformation module. The purpose is to convert discrete enterprise business, financial and operational data into continuous and semantically rich vector representations, providing a higher information density feature basis for subsequent multi-task processing. The deep feature representation refers to a low-dimensional dense vector obtained after processing by the feature encoding layer. This vector is a highly abstract and distributed representation of the original enterprise information. Each numerical value in the vector does not directly correspond to a specific business indicator, but collectively captures the intrinsic characteristics and potential patterns of the enterprise.

[0052] S32: The multi-task learning layer generates the main task score and the auxiliary task score corresponding to each enterprise in the candidate enterprise set based on the deep feature representation through parallel main task networks and auxiliary task networks. In this embodiment, the multi-task learning layer refers to a parallel task processing framework component layer, which can be implemented based on a shared bottom layer network structure with parameter isolation, such as by setting independent fully connected sub-networks to process the main task and auxiliary task respectively. The purpose is to focus the main task network on the core target, while using the auxiliary task network to capture related dimensions, and to achieve knowledge transfer between tasks through feature sharing mechanism, avoiding feature overfitting caused by single task training. The main task network refers to the neural network branch responsible for the core task in the multi-task learning architecture. In this scheme, its task is to regress or predict the main task score of the enterprise, i.e., the comprehensive score reflecting the overall spillover tendency of the enterprise. The auxiliary task network refers to the neural network branch that is parallel to the main task network but is responsible for predicting other related tasks. Auxiliary tasks are usually subtasks or side manifestations of the main task, such as predicting the auxiliary task score of the enterprise in specific dimensions such as "enterprise strength" and "spillover driving". Collaborative learning refers to a learning mechanism in which the main task and the auxiliary task promote and optimize each other during training. This collaboration is achieved by combining the losses of each task (such as main task regression loss + auxiliary task regression loss) in the loss function and updating the parameters of the shared layer and task-specific layer simultaneously through backpropagation.

[0053] S33: The fusion output layer performs weighted fusion on the main task score and the auxiliary task score corresponding to each enterprise to generate a set of enterprise comprehensive scores.

[0054] In the embodiment, the fusion output layer refers to a decision component layer integrating multi-dimensional score results, which can be implemented by using a dynamic weight distribution mechanism or a preset priority rule, for example, by using an adjustable linear combination function or an attention weight module. The purpose is to dynamically balance the contribution degrees of the main task and the auxiliary task according to the investment strategy, so as to ensure that the comprehensive score highlights the core target weight and effectively absorbs the supplementary information of the auxiliary dimensions; the weighted fusion refers to a strategy of linearly or nonlinearly combining the main task score and the auxiliary task score according to certain weights in the fusion output layer to generate the final enterprise comprehensive score.

[0055] Specifically, after the multi-dimensional information of the enterprise is converted into a deep feature representation by the feature encoding layer, the deep feature representation is input into the parallel network structure of the multi-task learning layer as a shared input, so that the main task network and the auxiliary task network are associated in the feature extraction stage, thereby strengthening the semantic consistency between the tasks in the collaborative learning process. The main task score and the auxiliary task score output by the multi-task learning layer are then transmitted to the fusion output layer to generate the final enterprise comprehensive score set through the weighted fusion mechanism. This process forms a complete information processing chain from the original data to the comprehensive decision, effectively solves the dimensional fragmentation problem caused by single task scoring, and avoids the interference of feature redundancy on the scoring accuracy.

[0056] Exemplarily, as a specific implementation, the feature encoding layer uses a pre-trained Transformer encoder to extract features from the multi-dimensional information of the enterprise, and encodes the heterogeneous data such as the business registration information, the intellectual property data and the financial statements of the enterprise into a 128-dimensional deep feature vector; the main task network in the multi-task learning layer is configured as a three-layer fully connected network for predicting the investment potential of the enterprise, and the three auxiliary task networks are respectively configured as double-layer fully connected networks for evaluating the technology matching degree, the environmental adaptability and the rule fit degree; the fusion output layer dynamically weights and fuses the main task score and the auxiliary task score output by the multi-task learning layer through a learnable attention weight module, wherein the weight parameters are initialized according to the regional investment decision priority.

[0057] Through the above technical solution, the application effectively integrates the deep features and the multi-task association in the enterprise comprehensive scoring process, so that the scoring results can take into account both the core investment target and the auxiliary evaluation dimensions, thereby avoiding the one-sided problem caused by traditional single task scoring, and providing a more comprehensive and more adaptive enterprise screening basis for subsequent multi-objective optimization, thereby significantly improving the accuracy and adaptability of the enterprise screening decision.

[0058] In an embodiment, step S32 comprises: S321: inputting the deep feature representation into the main task network and each auxiliary task network to generate an initial main score and a plurality of initial auxiliary scores through single forward propagation; In the present embodiment, the deep feature representation refers to a high-dimensional vector generated after feature extraction and coding of enterprise multi-dimensional information by a feature coding layer, which can be implemented by a convolutional neural network, a recurrent neural network or a Transformer architecture, aiming to comprehensively capture semantic features and potential correlations in enterprise multi-dimensional information; the main task network refers to a neural network structure focusing on core evaluation targets, which can be implemented by a multi-layer perceptron or an attention mechanism, aiming to accurately quantify main value indicators of an enterprise; the auxiliary task network refers to a neural network structure parallel to the main task network, used for evaluating supplementary dimensions, which can be implemented by the same or different architecture as the main task network, aiming to provide multi-angle perspectives for enterprise value evaluation; the single forward propagation refers to a process in which the deep feature representation starts from the input layer of the neural network, sequentially passes through all hidden layers (in the present scheme, the parallel main task network and auxiliary task network), and finally reaches the output layer, and all outputs (i.e., initial main scores and initial auxiliary scores) are calculated at one time; the initial main scores and the initial auxiliary scores refer to raw scores directly output by the main task network and the auxiliary task network without consistency processing, wherein the initial main scores refer to preliminary prediction values for core targets, and the initial auxiliary scores refer to preliminary prediction values in specific dimensions.

[0059] S322: constructing a score consistency optimization function with the initial main scores and the initial auxiliary scores as inputs, whose optimization goal is to minimize the difference between the weighted sum of the initial main scores and the weighted sum of the initial auxiliary scores, and to minimize the variance between the initial auxiliary scores; In the present embodiment, the score consistency optimization function refers to a mathematical constraint model used for coordinating the consistency of main task and auxiliary task scores, which can be implemented by the Lagrange multiplier method or the alternating direction multiplier method, aiming to eliminate score conflicts between tasks, meet specific logical consistency relationships and ensure the inherent stability of the scoring system; in the score consistency optimization function, the weighted sum of the initial main scores can be regarded as a general estimate of the main task scores, while the weighted sum of the initial auxiliary scores is a general score synthesized from scores in each dimension according to importance (weight), and the difference between the weighted sum of the initial main scores and the weighted sum of the initial auxiliary scores is optimized, aiming to make the main task scores consistent with the overall evaluation of each dimension; the variance is used in statistics to measure the dispersion degree of a group of data, and in the present scheme, minimizing the variance between the initial auxiliary scores means reducing the difference between different auxiliary task scores, for example, reducing the difference between the "strength" score and the "drive" score.

[0060] S323: generating the main task scores and the auxiliary task scores by solving the score consistency optimization function.

[0061] In the embodiment, the solving refers to finding the optimized main task score and auxiliary task score by a corresponding mathematical optimization algorithm (such as gradient descent method, least square method, etc.) so that the value of the above-mentioned score consistency optimization function reaches the minimum.

[0062] Specifically, the scheme first inputs the deep feature representation into the main task network and each auxiliary task network, generates initial main scores and initial auxiliary scores through single forward propagation, and ensures that all tasks are evaluated based on a unified feature basis; then a score consistency optimization function is constructed, which aligns the weighted sum of the main task scores and the weighted sum of the auxiliary task scores by force, and suppresses the variance within the auxiliary task scores, effectively eliminating the systematic conflicts caused by the target differences between tasks; finally, the initial scores are mapped to the optimization solution space under the consistency constraint by solving the optimization function, and the main task scores and auxiliary task scores that retain task specificity and meet overall coordination are generated, thereby providing a reliable basis for subsequent enterprise comprehensive scoring.

[0063] Exemplarily, as a specific implementation, the feature encoding layer adopts a BERT model to encode the multi-dimensional information of the enterprise to generate a deep feature representation; the main task network adopts a three-layer fully connected neural network structure to evaluate the investment potential of the enterprise; each auxiliary task network adopts a two-layer fully connected neural network structure to evaluate the technical strength, market influence and environmental adaptability of the enterprise; and the score consistency optimization function adopts a gradient descent method for solving, so as to achieve the best consistency between the main task scores and the auxiliary task scores through iterative optimization.

[0064] Through the above technical solution, the application effectively solves the comprehensive score distortion problem caused by score conflicts in multi-task learning, so that the enterprise comprehensive score can accurately reflect the real value of the enterprise and provide a more reliable basis for investment decisions.

[0065] In an embodiment, step S40 comprises: S41: analyzing the optimization target information to extract N optimization sub-targets and priority weights corresponding to each optimization sub-target; In the embodiment, the analysis of the optimization target information refers to the structural analysis of the screening dimensions set for the investment promotion area, which can be achieved by using the natural language processing technology combined with the pre-defined semantic rule library. Through the keyword extraction and semantic association analysis of the investment promotion regulation text, the quantified indicators and relative importance of each sub-target are determined. The purpose is to establish a hierarchical screening target system to avoid interference of key dimensions by secondary factors. The optimization sub-target refers to the specific and quantifiable sub-targets decomposed from the macro investment promotion strategy (optimization target information). These optimization sub-targets collectively define the multiple attributes that a qualified investment promotion scheme should have, such as "maximizing enterprise quality", "maximizing investment scale", and "maximizing industrial chain synergy effect". The priority weight refers to the numerical value used to quantify the relative importance of each optimization sub-target. The higher the priority weight, the higher the priority of the optimization sub-target in the final decision. Generally, the priority weight is set by the decision maker according to the strategic focus of this investment promotion.

[0066] S42: Calculate the value distribution statistics of each optimization sub-target corresponding to the target function on the candidate enterprise set based on the enterprise comprehensive score set, and normalize each target function according to the value distribution statistics to generate dimensionless standardized target functions; In the embodiment, the optimization sub-target corresponding target function is a mathematical function used to map the potential investment promotion scheme (i.e. a combination of candidate enterprises) to a calculable numerical value, which quantifies the performance of the scheme in a certain optimization sub-target. The value distribution statistics refers to the statistical characteristics of all possible values of the target function on the candidate enterprise set, such as maximum, minimum, average, variance, quantile, etc. By quantitatively describing the distribution characteristics of the candidate enterprise actual score data, the purpose is to eliminate the dimensional differences between different optimization sub-targets, making the heterogeneous dimensions such as economic benefits and technical matching comparable. The normalization processing refers to a data preprocessing technique used to eliminate the influence of dimensional and scale differences between different target functions. The core is to map the original value of each target function to a unified, dimensionless interval according to the value distribution statistics. The standardized target function refers to the dimensionless and scale-unified target function obtained after normalization processing.

[0067] S43: Analyze the regional resource constraint information to obtain rigid constraints and soft constraints; In the embodiment, the rigid constraint condition refers to a restriction condition that must be strictly met in the investment decision-making, which defines the boundary of the feasible region of the solution, for example, "the total land demand must not exceed the local available land area" is a typical rigid constraint; the soft constraint condition refers to a restriction condition that is expected to be met as much as possible but is allowed to be violated within a certain degree, and the violation of the soft constraint condition will lead to the reduction of the solution quality, but the solution itself is still feasible; specifically, the distinction between the rigid constraint condition and the soft constraint condition refers to the elastic classification of the constraint conditions according to the regional resource endowment characteristics, which can be realized by using a rule engine combined with an expert knowledge base, and the hard conditions that must be met and the soft conditions that can be flexibly adjusted are classified and managed, and the purpose is to provide differentiated modeling methods for the constraint conditions, to ensure the basic feasibility of the solution while retaining reasonable flexibility.

[0068] S44: convert the rigid constraint condition into a hard constraint inequality set, construct a corresponding penalty function based on the soft constraint condition, and take the constructed penalty function as an N+1th optimization sub-objective; In the embodiment, the hard constraint inequality set refers to a set of mathematical inequalities translated from the rigid constraint condition, which collectively defines the feasible solution space of the multi-objective optimization problem, and is used to convert the restrictions in the business language into mathematical language that can be understood and processed by the optimization algorithm; the penalty function refers to a mathematical technique for handling soft constraints or optimization objectives, which can be realized in parallel by using an exponential decay function or a piecewise linear function, and the purpose is to convert the flexible constraint into an optimized mathematical expression to avoid the limitations of the solution caused by simple filtering, and the core is that when the solution violates the soft constraint, the penalty function calculates the corresponding penalty value, and the penalty value is added to the overall optimization objective, the greater the violation degree, the higher the penalty value, and thus the lower the probability of being selected by the algorithm.

[0069] S45: combine the standardized objective function and the penalty function based on the priority weight of each optimization sub-objective, and construct a multi-objective optimization function based on the combination result and the hard constraint inequality set.

[0070] Specifically, the scheme first performs semantic analysis on the optimization target information to extract N sub-targets and their priority weights, and determines the importance level of each screening dimension; then, based on the enterprise comprehensive score set, the value distribution statistics of the objective function are calculated, and the statistical characteristics of the real performance data of the candidate enterprises are normalized to eliminate the dimensional differences between heterogeneous targets; at the same time, the regional resource constraint information is analyzed twice, the rigid constraint is converted into a hard constraint inequality group to ensure the feasibility of the scheme, and the soft constraint is constructed as a penalty function and as the N+1th optimization sub-target; finally, based on the priority weights of each sub-target, the standardized objective function and the penalty function are combined by weighting, and the complete multi-objective optimization function is constructed by fusing the hard constraint condition; the above process solves the misalignment problem of the objective function through dimensional normalization, coordinates the rigid and flexible requirements through constraint grading processing, ensures the dominance of the core screening dimension through priority weight allocation, and forms a systematic optimization function construction mechanism.

[0071] Exemplarily, as a specific embodiment, when a high-tech industrial cluster is invested, the optimization target information includes three sub-targets of investment intensity, industrial synergy degree and environmental compliance degree, and the priority weights of the sub-targets are extracted from the analysis of the investment regulation and policy documents as 0.5, 0.3 and 0.2 respectively; based on the enterprise comprehensive score set of the candidate enterprise set, the value distribution statistics (mean 5 million yuan / mu, standard deviation 1 million yuan / mu) of the investment intensity objective function are calculated, and the Z-score method is used for normalization processing; the regional resource constraint information analysis shows that the land supply is a rigid constraint (hard constraint inequality group: enterprise land demand ≤ 500 mu), and the tax refund ratio is a soft constraint, and accordingly an exponential penalty function with tax deviation degree as input is constructed; finally, the three normalized standardized objective functions and the penalty function are combined by weight to form a multi-objective optimization function containing hard constraint conditions, which is used for subsequent algorithm solving.

[0072] Through the above technical scheme, the present application effectively solves the misalignment problem of the optimization result caused by the dimensional difference of the objective function, ensures the comparability of heterogeneous dimensions such as economic benefit and technical matching in the optimization process; by distinguishing rigid constraints and soft constraints and constructing corresponding mathematical expressions, the screening scheme can meet the hard conditions and maintain reasonable flexibility on the soft conditions; at the same time, based on the optimization function combination mechanism of priority weights, the core screening dimension such as industrial synergy degree is ensured to occupy a dominant position in the decision-making, reducing the possibility of interference of secondary factors with the scientificity of the investment strategy, so as to generate an enterprise screening strategy scheme that conforms to the regional resource endowment characteristics and has practicality.

[0073] In an embodiment, step S50 comprises: S51: Extract an optimization feature vector based on the multi-objective optimization function and the set of hard constraint inequalities, the optimization feature vector including optimization problem size, constraint tightness, objective function nonlinearity, and optimization objective correlation; In this embodiment, the optimization feature vector refers to a vector quantifying the structural features of the multi-objective optimization problem, and the optimization feature vector includes optimization problem size, constraint tightness, objective function nonlinearity, and optimization objective correlation. The optimization problem size generally refers to the number of decision variables, which in this scenario is the number of candidate enterprises (n). The larger the optimization problem size, the exponentially increasing search space of the problem and the sharply increasing difficulty of solving the problem, reflecting the complexity of the decision variables. Constraint tightness is used to measure the strictness of the constraint condition, which can be quantified by calculating the proportion of feasible solutions (solutions that satisfy all constraints) in the total solution space. The lower the proportion, the tighter the constraint, and the more difficult the problem is to solve. Constraint tightness reflects the boundary conditions of the feasible region. The objective function nonlinearity is used to measure whether the objective function is a linear function. If the objective function or the constraint condition contains a nonlinear term (such as the synergistic effect between enterprises is not simply additive), the problem is nonlinear, and the objective function nonlinearity describes the curvature variation characteristics of the function. Optimization objective correlation refers to the degree of mutual promotion, mutual conflict, or mutual independence between multiple optimization objectives. For example, enterprise quality and investment size may be negatively correlated, i.e., high-quality enterprises may also be small in size. Optimization objective correlation is used to measure the degree of conflict between different optimization objectives.

[0074] S52: Perform similarity matching of the optimization feature vector with historical cases in the pre-constructed algorithm configuration library, the historical cases each containing a historical optimization feature vector and a corresponding preferred algorithm type and parameter configuration; In this embodiment, the algorithm configuration library refers to a knowledge base that stores the correspondence between historical problem features and successful algorithm configurations. Each record is a historical case, containing the problem features encountered in the past, the algorithm used and its parameters, and the performance of the configuration on the problem. Similarity matching refers to the process of numerically comparing the current optimization feature vector with the feature vectors of historical cases. Euclidean distance or cosine similarity algorithms can be used to achieve this. The purpose is to quickly locate historical cases that are highly similar to the current problem.

[0075] S53: Determine a target historical case based on the similarity matching result, and use the preferred algorithm type corresponding to the target historical case as the basic algorithm framework; In the embodiment, the determination of the target historical case refers to an operation of selecting an optimal matching historical case according to the similarity matching result, which can be realized by selecting the case with the highest similarity, and the purpose is to inherit the verified algorithm type as a starting point for solving; the basic algorithm framework refers to the algorithm type most similar to the current problem determined through similarity matching, which provides a basic paradigm and process for solving the problem, but the specific parameters need to be adjusted.

[0076] S54: Calculate the feature difference vector of the optimization feature vector and the historical optimization feature vector of the target historical case, and adjust the parameter configuration of the basic algorithm framework based on the preset parameter adjustment strategy and the feature difference vector.

[0077] In the embodiment, the feature difference vector refers to the specific difference vector of the current optimization feature vector and the historical optimization feature vector in each feature dimension, which can be realized by vector subtraction operation, and the purpose is to provide a quantitative basis for parameter adjustment; the parameter adjustment strategy refers to a predefined rule set or mapping function set for dynamically calibrating algorithm parameters based on the feature difference vector, which can be realized by converting the difference vector into a parameter change amount by using a preset mapping function, indicating how to modify the parameters of the basic algorithm framework according to the feature difference vector, for example, the parameter adjustment strategy can be: "if the current problem size is 50% larger than the historical case, then increase the population size parameter by 50%".

[0078] Specifically, the present scheme first extracts the optimization feature vector based on the multi-objective optimization function and the set of hard constraint inequalities, quantifies the key structural features from the nature of the problem, and avoids subjective interference of artificial experience; then, the optimization feature vector is matched with the historical cases in the algorithm configuration library for similarity, and the accumulated historical experience is used to quickly locate the highly similar problem scene; then, the target historical case is determined according to the matching result, and the preferred algorithm type is used as the basic algorithm framework to ensure the scientificity of the starting point for solving; finally, the feature difference vector is calculated and the parameters are adjusted based on the preset strategy, and the dynamic calibration is performed for the subtle differences; the above complete process from feature extraction, case matching, framework inheritance to parameter fine-tuning forms a closed-loop adaptive optimization mechanism, so that the algorithm selection and parameter configuration can closely match the characteristics of the current problem, significantly improving the efficiency and robustness of multi-objective optimization solving.

[0079] Exemplarily, as a preferred embodiment, in the screening of new energy vehicle field enterprises, an optimization feature vector is extracted based on a multi-objective optimization function and a set of hard constraint inequalities, wherein the problem size represents the order of magnitude of the decision variables, the constraint tightness reflects the boundary conditions of the feasible region, the nonlinearity of the objective function describes the curvature variation of the function, and the optimization target correlation measures the conflict degree between the targets; the feature vector is matched with the historical cases in the algorithm configuration library for similarity, and it is found that the similarity with a certain historical case (related to the screening of battery manufacturing enterprises) reaches 90%; according to the matching result, the historical case is determined as the target historical case, and the NSGA-II algorithm used by the historical case is used as the basic algorithm framework; after calculating the feature difference vector, based on the preset parameter adjustment strategy, the crossover probability is fine-tuned from 0.85 to 0.88, and the mutation probability is adjusted from 0.1 to 0.12, to adapt to the specific characteristics of the current problem.

[0080] Through the above scheme, the automatic matching and parameter adjustment of the multi-objective optimization algorithm are realized, the inefficient process of manually selecting the algorithm and adjusting the parameters is effectively avoided, the algorithm screening period is significantly shortened, the accuracy and convergence efficiency of the optimization solution are improved, and the quality of the enterprise screening strategy scheme set is improved.

[0081] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0082] In an embodiment, a target enterprise screening system for investment is provided, which is fused with constraint conditions and multi-objective optimization. The target enterprise screening system for investment fused with constraint conditions and multi-objective optimization corresponds to the target enterprise screening method for investment fused with constraint conditions and multi-objective optimization in the above embodiment. The target enterprise screening system for investment fused with constraint conditions and multi-objective optimization comprises: A screening module is configured to obtain target industry field information, screen a pre-constructed enterprise database based on the target industry field information, and generate an initial target enterprise pool according to a screening result; A first input module is configured to obtain regional resource constraint information and enterprise multi-dimensional information associated with the initial target enterprise pool, and input the enterprise multi-dimensional information and the regional resource constraint information into a pre-trained constraint discrimination model to obtain a candidate enterprise set; A second input module is configured to input enterprise multi-dimensional information associated with the candidate enterprise set into a pre-trained inclination evaluation model to obtain an enterprise comprehensive score set; An optimization function construction module is configured to obtain optimization target information, and construct a multi-objective optimization function based on the enterprise comprehensive score set and the optimization target information; The optimization algorithm adjustment module is configured to obtain optimization condition information, match a corresponding multi-objective optimization algorithm from a pre-constructed algorithm configuration library based on the optimization condition information and the multi-objective optimization function, and adjust parameters of the multi-objective optimization algorithm. The decision generation module is configured to solve the multi-objective optimization function by using the adjusted multi-objective optimization algorithm, generate a set of enterprise screening strategy schemes, and generate an enterprise screening decision report based on the set of enterprise screening strategy schemes.

[0083] Preferably, the screening module comprises: The feature vector generation submodule is configured to perform natural language processing and semantic expansion on the target industry field information, and generate an industry feature vector. The portrait feature construction submodule is configured to extract basic information data of each enterprise from an enterprise database, and construct an enterprise portrait feature set of each enterprise based on a pre-defined feature engineering rule library. The initial weight calculation submodule is configured to extract enterprise meta-features based on the enterprise portrait feature set, and calculate initial weights of the enterprise portrait feature set based on the enterprise meta-features. The similarity calculation submodule is configured to perform similarity calculation on the enterprise portrait feature set and the industry feature vector, sort each enterprise according to the similarity calculation result, and generate an initial target enterprise pool.

[0084] The specific limitations of the investment target enterprise screening system that integrates constraint conditions and multi-objective optimization can be seen in the limitations of the investment target enterprise screening method that integrates constraint conditions and multi-objective optimization described above, and will not be repeated here. Each module in the investment target enterprise screening system that integrates constraint conditions and multi-objective optimization can be realized by software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0085] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for selecting target enterprises for investment promotion that integrates constraints and multi-objective optimization, characterized in that, Including the following steps: Acquire target industry sector information, filter the pre-built enterprise database based on the target industry sector information, and generate an initial target enterprise pool based on the filtering results; Obtain regional resource constraint information and multi-dimensional information of enterprises associated with the initial target enterprise pool, and input the multi-dimensional information of enterprises and regional resource constraint information into a pre-trained constraint discrimination model to obtain a candidate enterprise set; The multidimensional information of the candidate enterprise set is input into a pre-trained propensity evaluation model to obtain a comprehensive enterprise score set. Obtain optimization target information and construct a multi-objective optimization function based on the enterprise comprehensive score set and optimization target information; Obtain optimization condition information, match the corresponding multi-objective optimization algorithm from the pre-built algorithm configuration library based on the optimization condition information and the multi-objective optimization function, and adjust the parameters of the multi-objective optimization algorithm; The multi-objective optimization function is solved by the adjusted multi-objective optimization algorithm to generate a set of enterprise screening strategy schemes, and an enterprise screening decision report is generated based on the set of enterprise screening strategy schemes.

2. The method for selecting target enterprises for investment promotion that integrates constraints and multi-objective optimization according to claim 1, characterized in that: The steps of acquiring target industry sector information, filtering the pre-built enterprise database based on the target industry sector information, and generating an initial target enterprise pool based on the filtering results include the following steps: Natural language processing and semantic expansion are performed on information from the target industry sector to generate industry feature vectors; Extract basic information data of each enterprise from the enterprise database, and construct enterprise profile feature sets for each enterprise based on a predefined feature engineering rule base; Enterprise meta-features are extracted based on the enterprise profile feature set, and the initial weights of the enterprise profile feature set are calculated based on the enterprise meta-features. The similarity between the enterprise profile feature set and the industry feature vector is calculated. Based on the similarity calculation results, the enterprises are ranked and an initial target enterprise pool is generated.

3. The method for selecting target enterprises for investment promotion that integrates constraints and multi-objective optimization according to claim 1, characterized in that: The constraint discrimination model includes a graph construction layer, a graph convolutional layer, and a readout layer. The step of obtaining regional resource constraint information and multi-dimensional information of enterprises associated with the initial target enterprise pool, and inputting the multi-dimensional information of enterprises and regional resource constraint information into the pre-trained constraint discrimination model to obtain a candidate enterprise set, includes the following steps: The graph construction layer constructs an enterprise relationship information graph based on multi-dimensional enterprise information, where the nodes of the enterprise relationship information graph are enterprises, and the relationships are edges; Graph convolutional layers transmit information and update node features in the enterprise association information graph, so that the node features of each node in the enterprise association information graph include the multi-hop neighborhood structure information of the enterprise corresponding to that node. The readout layer fuses regional resource constraint information with node features in the enterprise association information graph, performs global pooling and node-level discrimination on the fused features, and generates a candidate enterprise set based on the node-level discrimination results.

4. The method for selecting target enterprises for investment promotion that integrates constraints and multi-objective optimization according to claim 3, characterized in that: The graph convolutional layer, through information transfer and node feature updating of the enterprise association information graph, ensures that the node features of each node in the enterprise association information graph include the multi-hop neighborhood structure information of the enterprise corresponding to that node. The steps include: (a) Generate the message vector for each edge in the enterprise association information graph; (b) Aggregate the message vectors corresponding to the edges pointing to the nodes in the enterprise association information graph, and generate the aggregated adjacent message vectors for each node. (c) Fuse the aggregated adjacent message vectors of all nodes in the enterprise association information graph with their current feature representations to generate updated feature representations for each node and replace their current feature representations. When the fusion is performed for the first time, the current feature representation of the node is the enterprise multidimensional information. (d) Using steps (a), (b), and (c) as one iteration, repeat the iteration until the termination condition is met, so that the node features of each node in the enterprise association information graph include the multi-hop neighborhood structure information of the enterprise corresponding to that node, and the termination condition includes the number of iterations reaching the preset number of iterations K.

5. The method for selecting target enterprises for investment promotion that integrates constraints and multi-objective optimization according to claim 1, characterized in that: The propensity assessment model includes a feature encoding layer, a multi-task learning layer, and a fusion output layer. The step of inputting the multi-dimensional information of the candidate enterprise set associated with the pre-trained propensity assessment model to obtain the comprehensive enterprise score set includes the following steps: The feature encoding layer extracts and encodes features from the multidimensional information of enterprises, generating deep feature representations for each enterprise in the candidate enterprise set. The multi-task learning layer is based on deep feature representation and performs collaborative learning through a parallel main task network and several auxiliary task networks to generate the main task score and auxiliary task score for each enterprise in the candidate enterprise set. The fusion output layer performs weighted fusion of the main task scores and auxiliary task scores for each enterprise to generate a comprehensive enterprise score set.

6. The method for selecting target enterprises for investment promotion that integrates constraints and multi-objective optimization according to claim 5, characterized in that: The multi-task learning layer, based on deep feature representation, generates the main task score and auxiliary task score for each enterprise in the candidate enterprise set through collaborative learning via a parallel main task network and several auxiliary task networks. The steps include: The deep feature representation is input into the main task network and each auxiliary task network. Through a single forward propagation, an initial main score and several initial auxiliary scores are generated. Construct a rating consistency optimization function with inputs of initial main ratings and initial auxiliary ratings. Its optimization objective is to minimize the difference between the weighted sum of the initial main ratings and the weighted sum of the initial auxiliary ratings, while minimizing the variance among the initial auxiliary ratings. By solving the scoring consistency optimization function, the main task score and the auxiliary task score are generated.

7. The method for selecting target enterprises for investment promotion that integrates constraints and multi-objective optimization according to claim 1, characterized in that: The step of obtaining optimization target information and constructing a multi-objective optimization function based on the enterprise comprehensive score set and the optimization target information includes the following steps: The optimization objective information is analyzed to extract N optimization sub-objectives and the priority weights of each optimization sub-objective; Based on the comprehensive enterprise score set, the distribution statistics of the objective function corresponding to each optimization sub-objective on the candidate enterprise set are calculated, and each objective function is normalized according to the distribution statistics to generate a dimensionless standardized objective function. The regional resource constraint information is analyzed to obtain rigid constraints and soft constraints. The rigid constraints are transformed into a set of hard constraint inequalities. Based on the soft constraints, the corresponding penalty function is constructed, and the constructed penalty function is used as the N+1th optimization sub-objective. Based on the priority weights of each optimization sub-objective, the standardized objective function and penalty function are combined, and a multi-objective optimization function is constructed based on the combination result and a set of hard constraint inequalities.

8. The method for selecting target enterprises for investment promotion that integrates constraints and multi-objective optimization according to claim 7, characterized in that: The steps of obtaining optimization condition information, matching the corresponding multi-objective optimization algorithm from a pre-built algorithm configuration library based on the optimization condition information and the multi-objective optimization function, and adjusting the parameters of the multi-objective optimization algorithm include the following steps: The optimization feature vector is extracted based on a multi-objective optimization function and a set of hard-constraint inequalities. The optimization feature vector includes the optimization problem size, constraint tightness, nonlinearity of the objective function, and correlation of the optimization objective. The optimized feature vectors are matched with historical cases in a pre-built algorithm configuration library based on similarity. Each historical case contains historical optimized feature vectors and corresponding preferred algorithm types and their parameter configurations. The target historical cases are determined based on the similarity matching results, and the preferred algorithm type corresponding to the target historical cases is used as the basic algorithm framework. The algorithm calculates the feature difference vector between the optimized feature vector and the historical optimized feature vector of the target historical case, and adjusts the parameter configuration of the basic algorithm framework based on the preset parameter adjustment strategy and the feature difference vector.

9. A system for selecting target enterprises for investment promotion that integrates constraints and multi-objective optimization, characterized in that, include: The filtering module is used to obtain target industry information, filter the pre-built enterprise database based on the target industry information, and generate an initial target enterprise pool based on the filtering results. The first input module is used to obtain regional resource constraint information and multi-dimensional information of enterprises associated with the initial target enterprise pool, and input the multi-dimensional information of enterprises and regional resource constraint information into the pre-trained constraint discrimination model to obtain a candidate enterprise set. The second input module is used to input the multi-dimensional information of the candidate enterprise set into the pre-trained propensity evaluation model to obtain the comprehensive enterprise score set. The optimization function construction module is used to obtain optimization target information and construct multi-objective optimization functions based on the enterprise comprehensive score set and optimization target information; The optimization algorithm adjustment module is used to obtain optimization condition information, match the corresponding multi-objective optimization algorithm from the pre-built algorithm configuration library based on the optimization condition information and the multi-objective optimization function, and adjust the parameters of the multi-objective optimization algorithm. The decision generation module is used to solve the multi-objective optimization function using an adjusted multi-objective optimization algorithm, generate a set of enterprise screening strategy solutions, and generate an enterprise screening decision report based on the set of enterprise screening strategy solutions.

10. The investment target enterprise screening system integrating constraints and multi-objective optimization according to claim 9, characterized in that: The filtering module includes: The feature vector generation submodule is used to perform natural language processing and semantic expansion on target industry information to generate industry feature vectors. The profile feature construction submodule is used to extract basic information data of each enterprise from the enterprise database and construct the enterprise profile feature set of each enterprise based on the predefined feature engineering rule library; The initial weight calculation submodule is used to extract enterprise meta-features based on the enterprise profile feature set, and calculate the initial weights of the enterprise profile feature set based on the enterprise meta-features. The similarity calculation submodule is used to calculate the similarity between the enterprise profile feature set and the industry feature vector, sort the enterprises according to the similarity calculation results, and generate an initial target enterprise pool.

Images