Intelligent prediction recommendation method and system based on full-link data of supply chain

Abstract

The invention relates to the field of data processing, and provides an intelligent prediction recommendation method and system based on full-link data of a supply chain. The method comprises the steps of collecting multi-source heterogeneous supply chain data and performing standardization processing to obtain a standardized data set; according to the standardized data set, performing quantitative modeling on the supply chain state features to obtain a supply chain state feature vector; on the basis of the supply chain state feature vector, performing supply chain entity relationship modeling through a heterogeneous graph neural network to obtain a graph embedding vector; according to the graph embedding vector, performing dynamic generation on a candidate set through a multi-objective optimization algorithm to obtain a dynamic candidate set; and performing personalized matching of user requirements based on the dynamic candidate set to obtain a recommendation result. According to the invention, intelligent recommendation of the supply chain is realized, the decision-making efficiency is improved, and the operation cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an intelligent prediction recommendation method and system based on supply chain full-link data. BACKGROUND

[0002] Existing supply chain recommendation systems are mainly based on traditional collaborative filtering, content filtering or hybrid recommendation algorithms, which analyze historical transaction data, user behavior data and commodity attribute information to make recommendations. These systems usually use a relational database to store structured data and use machine learning algorithms such as support vector machines, decision trees or neural networks for feature learning and prediction modeling.

[0003] However, the prior art has the following disadvantages: first, traditional recommendation algorithms cannot effectively handle the complex multi-entity relationships and heterogeneous data fusion problems in the supply chain, resulting in a lack of global optimization perspective in the recommendation results; second, traditional methods often ignore the deep semantic relationships and path dependence characteristics between supply chain entities when modeling features, and cannot capture the implicit association patterns in the supply chain network; in addition, the existing technology lacks a unified mathematical modeling framework for multi-objective optimization, making it difficult to achieve a dynamic balance between efficiency, cost, satisfaction and other dimensions, resulting in insufficient feasibility and effectiveness of the recommendation results in practical applications. SUMMARY

[0004] The present application provides an intelligent prediction recommendation method and system based on supply chain full-link data to solve the defects of the prior art.

[0005] The present application provides an intelligent prediction recommendation method based on supply chain full-link data, comprising: S1: Collecting multi-source heterogeneous supply chain data and performing standardization processing to obtain a standardized data set; S2: Quantitative modeling of supply chain state characteristics based on the standardized data set to obtain a supply chain state feature vector; S3: Supply chain entity relationship modeling based on the supply chain state feature vector through a heterogeneous graph neural network to obtain a graph embedding vector; S4: Dynamic generation of a candidate set based on the graph embedding vector through a multi-objective optimization algorithm to obtain a dynamic candidate set; S5: Individualized matching of user demand based on the dynamic candidate set to obtain a recommendation result.

[0006] According to the intelligent prediction recommendation method based on supply chain full-link data provided by the present application, step S1 further comprises: S11: Collecting supply chain data to obtain raw data; the supply chain raw data includes: supplier raw data, logistics raw data, merchant raw data, and user raw data; S12: The original data is de-duplicated to obtain cleaned data; S13: The cleaned data is format-converted according to a unified data dictionary to obtain a standardized data set.

[0007] According to the intelligent prediction recommendation method based on supply chain full-link data provided by the application, step S2 further comprises: S21: From the standardized data set, the supply capacity dimension is extracted through a feature engineering framework to obtain a supply capacity feature; from the standardized data set, the demand fluctuation dimension is identified through time series analysis to obtain a demand fluctuation feature; and from the standardized data set, the market trend dimension is mined through a market analysis algorithm to obtain a market trend feature; S22: The supply capacity feature, the demand fluctuation feature and the market trend feature are weight-assigned through a random forest algorithm, and feature fusion is performed according to the assigned feature weight matrix to obtain a fused feature; S23: The supply chain health degree is quantitatively evaluated according to the fused feature through a fuzzy comprehensive evaluation method to obtain the supply chain state feature vector.

[0008] According to the intelligent prediction recommendation method based on supply chain full-link data provided by the application, step S3 further comprises: S31: The supply chain entity node is modeled to obtain a supply chain heterogeneous graph; wherein the node of the supply chain heterogeneous graph represents a supply chain entity, and the edge of the supply chain heterogeneous graph represents a supply chain relationship; S32: The supply chain heterogeneous graph is semantically subgraph-split to obtain a semantic subgraph set; S33: The node features in the semantic subgraph set are projected through a multi-layer perception to obtain a unified latent space representation; S34: According to the unified latent space representation, the isomorphic neighbors and the heterogeneous neighbors are respectively aggregated by introducing a multi-head attention of a meta-path weight matrix to obtain neighbor aggregation features; S35: The neighbor aggregation features are fused through node-level attention and semantic-level attention to obtain the graph embedding vector.

[0009] According to the intelligent prediction recommendation method based on supply chain full-link data provided by the application, step S34 further comprises: S341: Relationship paths are explored based on the supply chain heterogeneous graph to obtain a meta-path set; S342: The meta-path set is subjected to path representation learning to obtain a meta-path embedding vector; S343: The importance of the plurality of meta-paths is evaluated according to the meta-path embedding vector through similarity calculation, and a meta-path weight matrix is obtained; S344: The meta-path weight matrix is integrated into multi-head attention to obtain fusion attention, and the same structure neighbor and the heterogeneous neighbor are aggregated through the fusion attention to obtain neighbor aggregation features.

[0010] According to the intelligent prediction recommendation method based on supply chain full link data provided by the application, the expression of the fusion attention in step S344 is: ; Wherein, is a meta-path index, is an other path index, is a meta-path set, is a multi-head attention index, is the weight coefficient of the meta-path under the first head attention, is the query vector of the meta-path in the first head attention mechanism, is the key vector of the meta-path in the first head attention mechanism, is the position encoding of the meta-path under the first head attention, is the query vector of the other path in the first head attention mechanism, is the key vector of the other path in the first head attention mechanism, is the position encoding of the other path under the first head attention.

[0011] According to the intelligent prediction recommendation method based on supply chain full link data provided by the application, step S4 further comprises: S41: An optimization objective function is constructed based on the graph embedding vector; S42: The optimization objective function is globally searched through a genetic algorithm to obtain a candidate solution set; S43: The candidate solution set is verified for feasibility through constraint condition filtering to obtain the dynamic candidate set.

[0012] ​​​​​​​​According to the intelligent prediction and recommendation method based on end-to-end supply chain data provided by the present invention, the expression of the optimization objective function in step S41 is: ; in, For the efficiency objective function, These are the weighting coefficients corresponding to the efficiency objective function. Let the satisfaction objective function be... These are the weight coefficients corresponding to the satisfaction objective function. The objective function is cost. These are the weighting coefficients corresponding to the cost objective function; ; in, Indicates a time period. The total number of time periods. For the first Market demand over a specific period For the first Supply capacity for a given period of time. This is the demand fluctuation attenuation coefficient. This is the starting point of the current time window; ; in, For user index value, Total number of users For the first The actual user needs fulfillment rate For the first The potential needs of individual users For the first User preference weights for each user Indicates a point in time. For the first User behavior time points for each user The standard deviation of time decay; ; in, For cost item index value, This represents the total number of cost items. For the first Cost coefficient for each cost item For the first Decision variables for each cost item As an uncertainty penalty factor, The penalty coefficient for deviation from the benchmark value. The baseline value is obtained by fitting historical data.

[0013] According to the intelligent prediction recommendation method based on supply chain full-link data provided by the application, step S5 further comprises: S51: based on the dynamic candidate set, screening the supply chain layer candidate, obtaining the supply chain layer recommendation set; filtering the merchant layer candidate, obtaining the merchant layer recommendation set; personalized screening the user layer candidate, obtaining the user layer recommendation set; S52: input the supply chain layer recommendation set, the merchant layer recommendation set and the user layer recommendation set into the deep learning recommendation model for multi-task prediction, and obtain a prediction score matrix; S53: according to the prediction score matrix, the recommended candidate is sorted and output by the sorting algorithm, and the recommended result is obtained.

[0014] The application also provides an intelligent prediction recommendation system based on supply chain full-link data, comprising: The acquisition module is used for acquiring multi-source heterogeneous supply chain data and performing standardization processing to obtain a standardized data set; The quantification module is used for quantifying modeling the supply chain state characteristics according to the standardized data set to obtain a supply chain state characteristic vector; The modeling module is used for modeling the supply chain entity relationship based on the supply chain state characteristic vector through the heterogeneous graph neural network to obtain a graph embedding vector; The generation module is used for dynamically generating a candidate set according to the graph embedding vector through a multi-objective optimization algorithm to obtain a dynamic candidate set; The recommendation module is used for personalized matching of user demand based on the dynamic candidate set to obtain a recommended result.

[0015] The application provides an intelligent prediction recommendation method and system based on supply chain full-link data, which can effectively solve the problems of data islands and inconsistent formats in traditional supply chain systems by collecting multi-source heterogeneous supply chain data and performing standardized processing, laying a solid data foundation for subsequent intelligent analysis, enabling the system to comprehensively perceive the running state of the supply chain full-link, improving data utilization efficiency and decision-making accuracy. The technical features of the supply chain state feature quantization modeling based on the standardized data set, especially the multi-dimensional feature fusion method of extracting supply capacity features, time series analysis to identify demand fluctuation features, and market analysis algorithm to mine market trend features through the feature engineering framework, enable the system to comprehensively depict the complex state of the supply chain from multiple dimensions, significantly improving the completeness and accuracy of feature expression compared to traditional single-dimensional feature extraction methods, providing a more rich information foundation for accurate prediction and recommendation. The application of heterogeneous graph neural network in supply chain entity relationship modeling can effectively capture complex entity relationships and semantic information in the supply chain network by performing semantic subgraph splitting on the supply chain heterogeneous graph and using multi-layer perceptron for feature projection, especially introducing the multi-head attention mechanism of the meta-path weight matrix, enabling the system to adaptively identify and strengthen important relationship paths, significantly improving the accuracy and generalization ability of relationship modeling compared to traditional graph structure analysis methods, thereby providing more accurate graph embedding representation for subsequent intelligent decision-making. The application of multi-objective optimization algorithm can find the Pareto optimal solution set under complex constraints by constructing a comprehensive objective function containing efficiency, satisfaction, and cost dimensions and combining genetic algorithm for global search, significantly improving the global optimality and practicality of the solution compared to traditional single-objective optimization methods, especially in balancing and adapting to the dynamic changes of multiple interests in the supply chain. The personalized matching recommendation technology uses a hierarchical recommendation architecture design to perform targeted screening and filtering at the supply chain layer, merchant layer, and user layer, and uses a deep learning recommendation model for multi-task prediction, enabling the recommendation results to meet the individual needs of different levels at the same time, significantly improving the accuracy of the recommendation and user satisfaction compared to traditional unified recommendation methods.

[0016] Overall, the application applies artificial intelligence algorithms in the field of supply chain management, combining heterogeneous graph neural networks, multi-objective optimization algorithms, and deep learning recommendation models to make key contributions, not only improving the intelligence level and decision-making efficiency of the system, but also significantly improving the stability and sustainability of supply chain operations, providing strong technical support for the digital transformation and intelligent upgrading of supply chains. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 A flowchart of an intelligent prediction recommendation method based on supply chain full-link data provided by the present application is shown. Figure 2 A structural diagram of an intelligent prediction recommendation system based on supply chain full-link data provided by the present application is shown. DETAILED DESCRIPTION

[0019] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments. They should not be understood as limiting the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application. In the description of the present application, it should be understood that the terms used are only for the purpose of description, and should not be understood as indicating or implying relative importance.

[0020] As shown in Figure 1 The present application provides an intelligent prediction recommendation method based on supply chain full-link data, which comprises: S1: Collecting multi-source heterogeneous supply chain data and performing standardization processing to obtain a standardized data set.

[0021] The present application first collects multi-source heterogeneous supply chain data and performs standardization processing. Through systematic data collection, cleaning and standardization process, the heterogeneous data scattered in each link of the supply chain is converted into a standardized data set in a unified format, laying a solid data foundation for subsequent feature modeling and intelligent recommendation.

[0022] S1 further comprises: S11: Collecting supply chain data to obtain original data; the supply chain original data comprises: supplier original data, logistics original data, merchant original data and user original data.

[0023] Further, the application collects supply chain data to obtain raw data in the S11 step, and specifically extracts multi-dimensional data information from each participating entity in the entire supply chain. The supplier raw data includes supplier basic information, production capacity data, inventory level data, delivery history data, quality evaluation data, and other multi-dimensional information. The production capacity data records the upper limit of production capacity, the current production capacity utilization rate, the equipment state parameters, and other dynamic production indicators in each time period. The inventory level data reflects the real-time changes of raw material inventory, semi-finished product inventory, and finished product inventory. The delivery history data includes order completion time, delivery accuracy rate, transportation time, and other historical performance indicators. The quality evaluation data covers product qualification rate, customer satisfaction score, quality complaint record, and other quality-related information.

[0024] The logistics raw data covers transportation path information, warehouse management data, delivery time data, and logistics cost data, and other key logistics elements. The transportation path information records the starting point, the passing node, the destination, the transportation mode, the transportation distance, and other path-related parameters. The warehouse management data includes warehouse location coordinates, storage capacity, goods in and out of the warehouse time, inventory turnover rate, and other warehouse operation indicators. The delivery time data reflects order processing time, sorting and packaging time, last mile delivery time, and other time efficiency indicators. The logistics cost data covers transportation costs, warehouse costs, labor costs, fuel consumption, and other cost components.

[0025] The merchant raw data includes merchant basic archives, sales history records, inventory management information, and customer relationship data, and other merchant operation-related information. The merchant basic archives record the merchant number, the scope of operation, the business license information, and the scale of operation. The sales history records include transaction data such as product sales quantity, sales amount, sales time, and promotion activity information. The inventory management information reflects the inventory level, the replenishment frequency, and the inventory turnover of the merchant side. The customer relationship data covers customer purchase behavior, customer preference analysis, and customer satisfaction feedback.

[0026] The user raw data involves user basic information, behavior trajectory data, purchase history data, preference label data, and other user portrait-related information. The user basic information includes user identification, age group, gender, geographic location, income level, and other demographic characteristics. The behavior trajectory data records the user's browsing path, dwell time, click behavior, search keywords, and other behavior sequences on the platform. The purchase history data reflects the user's consumption patterns such as product category, purchase quantity, purchase frequency, and purchase time. The preference label data includes labeled information such as product category preference, brand preference, and price sensitivity generated through user behavior analysis.

[0027] S12: The raw data is processed to obtain cleaned data.

[0028] Further, in step S12, the original data is de-duplicated to obtain cleaned data, which aims to identify and remove duplicate records and redundant information in the data set through multi-level data de-duplication algorithm and data quality detection mechanism. The de-duplication process first adopts a fast de-duplication algorithm based on a hash function to quickly identify completely identical duplicate records by calculating the hash value of each data record. The hash function adopts MD5 algorithm to calculate the hash value of the key field of the data record, and generates a 128-bit hash value as the unique identifier of the record. When records with the same hash value are found, the system will further perform field-by-field comparison and verification to confirm the complete consistency of the records and remove duplicate records. For partially repeated or approximately repeated data records, an edit distance-based similarity matching algorithm is used for identification. The edit distance algorithm measures the similarity of two records by calculating the minimum number of editing operations required to convert one string to another. When the edit distance is less than a predetermined threshold, the two records are considered approximately repeated, and the system will select the record with higher quality according to data timestamp, data source weight and other factors. The data quality detection mechanism identifies and processes abnormal values, missing values and error values in the data through multi-dimensional data integrity check, data consistency verification, data accuracy evaluation and other means. The integrity check verifies the data integrity of the key fields, marks and processes records with missing key information, the consistency verification checks the logical consistency of cross-table associated data, finds and corrects logical contradictions between data, and the accuracy evaluation evaluates the accuracy of the data by comparing with historical data, business rule verification and other ways, and marks and processes obviously incorrect data records.

[0029] S13: Format conversion is performed on the cleaned data according to the unified data dictionary to obtain a standardized data set.

[0030] Further, in step S13, the application converts the heterogeneous data from different data sources into a unified standard format through predefined data dictionary specifications and format conversion rules. The unified data dictionary defines the standard field names, data types, value ranges, coding rules, and other specification requirements of the supply chain full-link data. The field names use a unified naming convention, the data types are strictly defined according to standard data types, including integer, floating-point, string, date and time, and Boolean basic data types, the value range defines the valid value interval of the field through minimum value, maximum value, and enumeration value, and the coding rule uniformly defines the coding standard and coding mapping relationship of the classification data. The format conversion process first standardizes the field name mapping, mapping various field names in the original data to the standard field names in the unified data dictionary. The mapping process uses a combination of rule-based field matching algorithm and machine learning semantic matching algorithm. The rule matching algorithm directly matches through a predefined field mapping table, and the semantic matching algorithm uses a word vector model to calculate the semantic similarity of the field name for intelligent matching. The data type conversion converts the original data into standard data types through a type conversion function. Integer conversion handles the conversion of numeric string to integer, floating-point conversion handles the conversion of decimal and scientific notation, string conversion handles the conversion of encoding format and special character processing, date and time conversion converts various date and time formats to standard format, and Boolean conversion converts various Boolean value representations to standard true / false format. Data coding standardization converts various coding forms of classification data into a unified coding standard. The coding conversion uses a coding mapping table for batch conversion, and a coding dictionary is established to record the correspondence between the original coding and the standard coding, facilitating subsequent data tracing and reverse conversion.

[0031] S2: quantitatively model the supply chain state characteristics according to the standardized data set, and obtain a supply chain state characteristic vector.

[0032] In step S2, the application quantitatively models the supply chain state characteristics according to the standardized data set. In this stage, through multidimensional feature engineering, time series analysis, market analysis and other technical means, the original data in the standardized data set is converted into a feature vector representation that can fully reflect the running state of the supply chain.

[0033] In step S2, the application quantitatively models the supply chain state characteristics according to the standardized data set. In this stage, through multidimensional feature engineering, time series analysis, market analysis and other technical means, the original data in the standardized data set is converted into a feature vector representation that can fully reflect the running state of the supply chain. S21: extract the supply capacity dimension from the standardized data set through a feature engineering framework to obtain a supply capacity feature; identify the demand fluctuation dimension from the standardized data set through time series analysis to obtain a demand fluctuation feature; and mine the market trend dimension from the standardized data set through a market analysis algorithm to obtain a market trend feature.

[0034] In the step S21, the supply capacity dimensions are extracted by the feature engineering framework to obtain supply capacity features. The feature engineering method based on statistical learning is used to extract quantitative indicators reflecting the production capacity, delivery capacity, quality control capacity and other dimensions of the supplier from the original data of the supplier. The feature engineering framework first performs statistical analysis on the historical production data of the supplier to calculate the production capacity utilization rate feature. The feature is measured by the ratio of actual output to theoretical maximum capacity, and the calculation formula is production capacity utilization rate equals to actual monthly output divided by theoretical monthly capacity. Then the on-time delivery rate feature is extracted. The feature is calculated by the ratio of the number of orders delivered on time within a certain time window to the total number of orders. At the same time, the quality pass rate feature is extracted. The feature is characterized by the ratio of the number of qualified products to the total number of products to reflect the quality control level of the supplier. The feature engineering framework also calculates the inventory turnover rate feature of the supplier, which reflects the efficiency of the supplier's inventory management. The feature is calculated by the ratio of annual total out-of-stock quantity to average inventory. In addition, the cost control ability feature of the supplier is extracted. The feature measures the cost control effect of the supplier by analyzing the unit production cost trend. For the demand fluctuation dimension, the demand data is deeply mined by time series analysis. The time series analysis uses ARIMA model to identify the trend, periodicity and randomness features in the demand data. The ARIMA model stands for autoregressive integrated moving average model. The model describes the dynamic characteristics of time series data through the combination of autoregressive term, difference term and moving average term. Through the analysis of historical demand data by ARIMA model, demand trend feature, demand seasonal feature and demand random fluctuation feature are extracted. The demand trend feature reflects the long-term growth or decline trend of demand, the demand seasonal feature captures the periodic change pattern of demand, and the demand random fluctuation feature describes the irregular change degree of demand. The time series analysis also uses sliding window method to calculate demand fluctuation intensity feature. The feature measures the fluctuation degree of demand by calculating the standard deviation of demand data within a fixed time window. The calculation formula is demand fluctuation intensity equals to the standard deviation of demand data within a time window divided by the average value of demand data in the same period. The market trend dimension is mined by market analysis algorithm to obtain market trend feature. The market analysis algorithm uses machine learning-based trend identification method to analyze market development trend combining external market data and internal transaction data. The market analysis algorithm first uses linear regression method to analyze market price trend. The trend direction and change speed of price change are identified by establishing a linear regression model of price and time. The expression of linear regression model is price equals to intercept plus slope multiplied by time plus error term, where the slope parameter reflects the strength and direction of price change trend.The market analysis algorithm also calculates the market demand growth rate feature using the moving average method, which identifies the growth or contraction trend of the market by comparing the market demand in different time periods. The calculation method is to subtract the market demand in the previous time period from the market demand in the current time period, and then divide by the market demand in the previous time period. In addition, the market analysis algorithm also extracts the market competition intensity feature, which measures the intensity of market competition by analyzing indicators such as the number of competitors, market share distribution, and price competition intensity in the market.

[0035] S22: The supply capacity feature, demand fluctuation feature, and market trend feature are assigned weights using the random forest algorithm, and the features are fused according to the assigned feature weight matrix to obtain the fused feature.

[0036] Further, the supply capacity feature, demand fluctuation feature, and market trend feature are assigned weights using the random forest algorithm in the S22 step, and the features are fused according to the assigned feature weight matrix to obtain the fused feature. The process uses a random forest-based feature importance evaluation method to determine the importance of different features in the supply chain state evaluation. The random forest algorithm improves the prediction accuracy by constructing multiple decision trees and voting or averaging their prediction results. In the feature weight assignment process, the random forest algorithm evaluates the importance of features by calculating the contribution of each feature in the decision tree splitting process.

[0037] Further, the present application calculates the importance scores of the supply capacity feature, demand fluctuation feature, and market trend feature using the random forest algorithm, and then normalizes these scores to form a feature weight matrix. The normalization process is achieved by dividing the importance score of each feature by the sum of all feature importance scores, ensuring that the sum of all feature weights is 1.

[0038] The feature fusion process uses a weighted average method to linearly combine features of different dimensions according to their weights. The dimensions include supply capacity features, demand fluctuation features, and market trend features. The feature fusion process also uses principal component analysis to reduce the dimensionality of the fused features. Principal component analysis achieves feature dimensionality reduction by finding the direction with the largest variance in the data. The reduced feature vector retains the main information of the original features while reducing the feature dimensionality.

[0039] S23: The supply chain health degree is quantitatively evaluated based on the fused feature using the fuzzy comprehensive evaluation method, and the supply chain state feature vector is obtained.

[0040] Further, the application obtains a supply chain state feature vector by quantitatively evaluating the supply chain health degree according to the fusion features in the S23 step through a fuzzy comprehensive evaluation method, and converts the fusion features into a health degree index capable of comprehensively reflecting the running state of the supply chain by using a comprehensive evaluation method based on fuzzy mathematics theory.

[0041] The fuzzy comprehensive evaluation method of the application first establishes an evaluation index system of the supply chain health degree, the index system including evaluation indexes of multiple dimensions such as supply stability, demand matching degree, cost control level, quality guarantee capability and risk resistance capability, each evaluation index corresponding to a fuzzy set, and the fuzzy set describing the membership degree of the index value to different health degree grades by a membership function. The membership function is represented by a trapezoidal fuzzy number, and the trapezoidal fuzzy number is defined by four parameters including a left boundary, a left core, a right core and a right boundary.

[0042] The fuzzy comprehensive evaluation method uses a weighted average operator to comprehensively evaluate each evaluation index, and after obtaining the comprehensive evaluation result of the supply chain health degree by the fuzzy comprehensive evaluation method, combines the result with the fusion feature vector to form a supply chain state feature vector, and the feature vector includes comprehensive information of the running state of the supply chain and the evaluation result of the health degree.

[0043] S3: Based on the supply chain state feature vector, a supply chain entity relationship model is established through a heterogeneous graph neural network to obtain a graph embedding vector.

[0044] The step S3 further includes: S31: A supply chain entity node is modeled to obtain a supply chain heterogeneous graph, wherein the node of the supply chain heterogeneous graph represents a supply chain entity, and the edge of the supply chain heterogeneous graph represents a supply chain relationship.

[0045] Furthermore, in step S31, the process of modeling supply chain entity nodes to obtain a heterogeneous supply chain graph involves abstracting real-world supply chain entities into nodes and edges in a graph structure. In the heterogeneous supply chain graph, nodes represent different types of entities in the supply chain, and edges represent relationships between entities. Node types include supplier entities, manufacturer entities, distributor entities, retailer entities, logistics service provider entities, and end-user entities. Each node carries a supply chain state feature vector obtained from step S2 as its initial feature representation. Edge types reflect the business relationships between entities, including supply relationships, procurement relationships, logistics relationships, and sales relationships. The construction process of the heterogeneous graph first requires identifying all entities participating in supply chain activities, and then determining the types of relationships between entities based on the actual business process. Supplier entity nodes contain feature information such as supply capacity, product quality, and delivery time; manufacturer entity nodes contain feature information such as production capacity, inventory level, and cost structure; and distributor entity nodes contain feature information such as distribution network coverage and inventory management capabilities. This feature information directly originates from the supply chain state feature vector obtained through quantitative modeling in step S2, ensuring consistency between the graph structure and the actual supply chain state.

[0046] S32: Perform semantic subgraph splitting on the heterogeneous supply chain graph to obtain a set of semantic subgraphs.

[0047] Furthermore, a semantic subgraph refers to a subgraph structure composed of nodes and edges with the same or similar semantic meanings in the original heterogeneous graph. The splitting process is based on the concept of meta-paths, which are path patterns connecting different types of nodes in a heterogeneous graph. The semantic subgraph splitting algorithm first identifies all possible meta-paths in the heterogeneous graph, and then decomposes the graph structure into multiple semantic subgraphs according to the semantic meaning of the meta-paths. Each semantic subgraph represents a specific business scenario or relationship pattern in the supply chain. The specific implementation process of semantic subgraph splitting includes meta-path enumeration, path instance extraction, and subgraph clustering steps. The meta-path enumeration process traverses all possible paths in the heterogeneous graph using a depth-first search algorithm and classifies them according to the sequence pattern of node types in the path. The path instance extraction process extracts specific path instances under each meta-path pattern to form the corresponding semantic subgraph. The subgraph clustering process uses graph structure similarity metrics to merge similar semantic subgraphs to form the final set of semantic subgraphs.

[0048] S33: Using a multilayer perceptron, the node features in the semantic subgraph set are projected to obtain a unified latent space representation.

[0049] Step S33 is a process of feature projection on node features in the semantic subgraph set by a multi-layer perception machine to obtain a unified latent space representation. It is a key link to realize the standardization of heterogeneous graph node features. The multi-layer perception machine is a kind of feedforward neural network composed of multiple fully connected layers, which can learn the nonlinear mapping relationship between input features and output features. In the supply chain heterogeneous graph, different types of nodes have feature vectors with different dimensions and meanings. Direct calculation will cause the problem of feature mismatch. The multi-layer perception machine maps the feature vectors of different types of nodes into a unified latent space through nonlinear transformation, so that the feature vectors of different types of nodes have the same dimension and comparability. The multi-layer perception machine is adjusted according to the feature complexity and computing resources. The training process optimizes the weight parameters through the back propagation algorithm, so that the projected feature vectors can retain the semantic information of the original features while having a unified representation format.

[0050] S34: According to the unified latent space representation, the isomorphic neighbors and the heterogeneous neighbors are aggregated respectively by introducing a multi-head attention of a meta-path weight matrix to obtain neighbor aggregation features.

[0051] Further, the isomorphic neighbors refer to neighbor nodes of the same type as the current node, and the heterogeneous neighbors refer to neighbor nodes different from the current node. The multi-head attention mechanism can simultaneously focus on multiple different aspects of information. The multi-head attention mechanism calculates different attention distributions in parallel through multiple attention heads, and then fuses the results. The meta-path weight matrix reflects the importance of different meta-paths in the modeling of supply chain relationships. By integrating the meta-path weight into the multi-head attention mechanism, the complex relationship patterns in the supply chain can be more accurately captured. The neighbor aggregation process first calculates the attention weight of each neighbor node to the current node, and then performs weighted average on the feature vectors of the neighbor nodes according to the attention weight to obtain the aggregated neighbor feature representation.

[0052] Step S34 further comprises: S341: Based on the supply chain heterogeneous graph, a set of meta-paths is obtained.

[0053] In step S341, the meta-path is a path pattern connecting different types of nodes in the heterogeneous graph, and each meta-path represents a specific business process or relationship pattern in the supply chain. The relationship path discovery process systematically searches all possible paths in the heterogeneous graph through a graph traversal algorithm and classifies and abstracts them according to the sequence pattern of node types in the path. The discovery process first performs a depth-first search starting from each node, records the sequence of node types passed during traversal, and stops searching when the path length reaches a preset threshold or meets the termination condition; then, it extracts patterns from all searched paths, and classifies paths with the same node type sequence into the same meta-path. The construction of the meta-path set needs to consider path length, path frequency, business semantics, etc., and the final meta-path set is determined through statistical analysis and expert knowledge.

[0054] S342: Perform path representation learning on the meta-path set to obtain meta-path embedding vectors.

[0055] Further, in step S342, the process of the present application for obtaining meta-path embedding vectors by performing path representation learning on the meta-path set aims to convert abstract path patterns into computable vector representations. Path representation learning is a technique that encodes path information in a graph structure into low-dimensional vectors, and generates vector representations of paths by learning the combination patterns of nodes and edges in the path. Meta-path embedding vectors can capture the structural and semantic information of the path, providing a basis for subsequent path importance evaluation. Specifically, the path representation learning process of the present application first maps each node type and edge type in the meta-path to an initial vector representation, then encodes the path sequence through a sequence encoder (such as a recurrent neural network or Transformer) to generate the embedding vector of the meta-path; the encoding process considers the order information of nodes and edges in the path, and maintains the structural characteristics of the path through position encoding and sequence modeling techniques; the training process optimizes the quality of the embedding vector through contrastive learning or reconstruction learning, so that similar meta-paths have similar vector representations, and different meta-paths have distinctive vector representations.

[0056] S343: Evaluate the importance of multiple meta-paths according to the meta-path embedding vectors through similarity calculation to obtain a meta-path weight matrix.

[0057] Further, in step S343, the application aims to obtain a meta-path weight matrix to reflect the importance of different meta-paths in a specific task or scenario. By assigning weights, important relationship patterns can be highlighted, and the influence of noise relationships can be suppressed. The importance evaluation process first calculates the similarity between meta-path embedding vectors, and in this embodiment, cosine similarity is used. Then, according to the similarity distribution and business prior knowledge, the importance weight of each meta-path is determined. The weight calculation process considers multiple factors, including path frequency, path length, node importance, edge importance, etc. The mathematical expression of the final meta-path weight matrix is: ; wherein, represents the weight value between meta-path and meta-path , and represent the embedding vectors of meta-path and meta-path , represents the similarity function, is the meta-path index, represents the size of the meta-path set.

[0058] S344: Integrate the meta-path weight matrix into multi-head attention to obtain a fusion attention. Through the fusion attention, the same and different neighbors are aggregated respectively to obtain neighbor aggregation features.

[0059] wherein, the expression of the fusion attention in step S344 is: ; wherein, is the meta-path index, is the other path index, is the meta-path set, is the multi-head attention index, is the weight coefficient of meta-path under the th attention, is the query vector of meta-path in the th attention mechanism, is the key vector of meta-path in the th attention mechanism, is the sequence length, is the position encoding of meta-path under the th attention, is the other path in the query vectors, for the first other paths under the first attention mechanism key vectors, for the first other paths under the first attention mechanism position encodings.

[0060] Further, step S344 integrates the meta-path weight matrix into the multi-head attention to obtain the fusion attention, and the process of aggregating the isomorphic neighbors and the heterogeneous neighbors respectively through the fusion attention to obtain the neighbor aggregation features is the core computing step of the heterogeneous graph neural network. The fusion attention mechanism can more accurately capture the complex relationship patterns in the supply chain by combining the multi-head attention and the meta-path weight information. The multi-head attention mechanism allows the model to focus on multiple different relationship aspects at the same time, and each attention head focuses on capturing a specific relationship pattern. The integration of the meta-path weight matrix enables the attention mechanism to adaptively adjust according to the importance of the path, improving the accuracy of relationship modeling. The calculation process of the fusion attention first allocates different meta-path weights to each attention head, then calculates the attention distribution between the query vectors, the key vectors and the value vectors, and finally aggregates the neighbor node features according to the attention weights. The isomorphic neighbor aggregation process focuses on the relationship patterns between nodes of the same type, and the heterogeneous neighbor aggregation process focuses on the relationship patterns between nodes of different types, and the two aggregation methods capture relationship information at different levels.

[0061] S35: fuse the neighbor aggregation features through the node-level attention and the semantic-level attention to obtain the graph embedding vector.

[0062] Further, step S35 fuses the neighbor aggregation features through the node-level attention and the semantic-level attention to obtain the graph embedding vector, which is the final output stage of the heterogeneous graph neural network. The node-level attention focuses on the importance of individual nodes in the graph structure, and the semantic-level attention focuses on the importance of different semantic relationships. The combination of the two-level attention mechanism can integrate graph structure information from different granularities to generate more comprehensive and accurate graph embedding vectors. The node-level attention assigns importance weights to each node by learning the relevance of node features and graph structure information, and the semantic-level attention assigns importance weights to each semantic relationship by learning the contribution of different semantic subgraphs. The fusion process integrates the results of the two-level attention through weighted averaging or a gating mechanism to generate the final graph embedding vector. The graph embedding vector contains local feature information of nodes, neighbor relationship information and global structure information, providing rich feature representation for subsequent multi-objective optimization and recommendation generation.

[0063] To better understand the isomorphic graph neural network modeling process of step S3 of the present application, a specific supply chain scenario is used as an example for illustration. Assume that there is an electronic product supply chain including chip suppliers A, circuit board manufacturers B, assembly factories C, brand manufacturers D, and retailers E. In step S31, when constructing the isomorphic graph, these entities are taken as nodes of different types, and the node features of chip supplier A include its monthly capacity of 100,000 pieces, yield rate of 98.5%, delivery period of 15 days, and other information, and the node features of circuit board manufacturer B include its monthly capacity of 80,000 pieces, inventory level of 5,000 pieces, cost of 50 yuan per piece, and other information. The relationships between entities include A supplying chips to B, B supplying circuit boards to C, C supplying assembled products to D, D supplying finished products to E, and other relationships, which constitute the edges of the isomorphic graph. In step S32, the isomorphic graph is split into semantic subgraphs such as a procurement subgraph, a production subgraph, and a sales subgraph according to business semantics, the procurement subgraph includes procurement relationships from suppliers to manufacturers, the production subgraph includes production process relationships within manufacturers, and the sales subgraph includes sales relationships from manufacturers to retailers. In step S33, the feature vectors of different types of nodes are mapped to a unified 128-dimensional latent space through a multi-layer perception, and the original feature vector of chip supplier A has a dimension of 32, and after transformation by the multi-layer perception, a 128-dimensional unified representation vector is obtained. In step S34, meta-paths such as “supplier-manufacturer-brand manufacturer” and “manufacturer-brand manufacturer-retailer” are identified, 64-dimensional embedding vectors of each meta-path are obtained through path representation learning, and then the cosine similarity between meta-paths is calculated to construct a weight matrix, which is integrated into an 8-head attention mechanism to aggregate the isomorphic neighbors (other suppliers) and the heterogeneous neighbors (manufacturer B) of chip supplier A, respectively, to obtain the aggregated neighbor feature vectors. In step S35, the importance weight of chip supplier A in the entire supply chain is learned to be 0.3 through node-level attention, and the importance weight of the procurement relationship is learned to be 0.4 through semantic-level attention, and finally a 128-dimensional graph embedding vector of chip supplier A is obtained through fusion, which comprehensively reflects the characteristics of chip supplier A, neighbor relationships, and position in the supply chain.

[0064] S4: dynamically generating a candidate set from the graph embedding vector through a multi-objective optimization algorithm to obtain a dynamic candidate set.

[0065] In step S4, the present application converts abstract graph structure information into specific recommendation schemes, specifically by constructing a comprehensive optimization objective function, using a genetic algorithm for global search, and filtering through constraint conditions to achieve dynamic generation of the candidate set.

[0066] In step S4, the present application converts abstract graph structure information into specific recommendation schemes, specifically by constructing a comprehensive optimization objective function, using a genetic algorithm for global search, and filtering through constraint conditions to achieve dynamic generation of the candidate set. S41: constructing an optimization objective function based on the graph embedding vector.

[0067] In step S41, the application constructs an optimization objective function based on graph embedding vectors. The optimization objective function established by the application considers the efficiency, satisfaction, and cost of supply chain operation in three dimensions. The graph embedding vectors serve as input data and contain high-dimensional feature representations of supply chain entities and their relationships. By incorporating these feature information into the construction process of the objective function, the optimization objective can accurately reflect the actual state and constraints of the supply chain.

[0068] The expression of the optimization objective function in step S41 is as follows: ; Wherein, is the efficiency objective function, is the weight coefficient corresponding to the efficiency objective function, is the satisfaction objective function, is the weight coefficient corresponding to the satisfaction objective function, is the cost objective function, is the weight coefficient corresponding to the cost objective function. ; Wherein, denotes a time period, is the total number of time periods, is the market demand in the th time period, is the supply capacity in the th time period, is the demand fluctuation decay coefficient, is the starting point of the current time window. ; Wherein, is the user index value, is the total number of users, is the actual demand satisfaction rate of the th user, is the potential demand of the th user, is the user preference weight of the th user, denotes a time point, is the user behavior time point of the th user, is the time decay standard deviation. ; Wherein, is the cost item index value, is the total number of cost items, is the cost of the a cost coefficient of the cost term, a decision variable of the first cost term, an uncertainty penalty factor, a penalty coefficient of deviating from the benchmark value, a benchmark value obtained by fitting historical data.

[0069] Further, the efficiency objective function expression reflects the matching efficiency of the supply chain by calculating the ratio of demand to supply in each time period, and an exponential decay function is used to simulate the influence of time on efficiency evaluation, that is, the farther away from the current time, the smaller the contribution of the time period to the overall efficiency; the satisfaction objective function calculates the satisfaction contribution of all users by weighted summation, and specifically simulates the time sensitivity of user behavior by using a Gaussian decay function, that is, the farther away from the current time, the smaller the influence of the user behavior on the satisfaction; the expression of the cost objective function includes a linear cost term and a quadratic penalty term, the linear cost term reflects the direct cost expenditure, and the quadratic penalty term is used to constrain the degree of deviation of the decision variable from the historical benchmark value, so as to avoid the risk caused by excessive deviation from historical experience.

[0070] S42: performing global search on the optimization objective function by using a genetic algorithm to obtain a candidate solution set.

[0071] In step S42, the genetic algorithm is used to search for the optimal solution. Specifically, first, an individual coding scheme is designed to encode the supply chain decision variable into a chromosome structure, the coding scheme adopts a real number coding method, and the initial population is generated by using a random sampling method to randomly generate individuals in the feasible region of the decision variable, forming an initial population , wherein is an individual representation vector.

[0072] The fitness function of the genetic algorithm is designed based on the value of the multi-objective optimization function. Since there are multiple conflicting objectives, the Pareto dominance relationship is used to evaluate individuals. The dominance relationship is that an individual dominates an individual if and only if the performance on all objectives is not worse than , and the performance on at least one objective is better than . The non-dominated sorting algorithm divides the individuals in the population into layers according to the Pareto dominance relationship, the first layer contains non-dominated individuals, the second layer contains individuals dominated by the first layer, and so on. The fitness value is determined comprehensively according to the dominance level and the crowding distance of the individual in the target space. The crowding distance measures the distribution density of the individual in the target space.

[0073] The selection operation adopts a tournament selection mechanism to randomly select The individuals are compared, and the individual with the highest fitness is selected as the parent. The crossover operation adopts a simulated binary crossover operator to recombine the genes of the selected parent individuals. The calculation process of the simulated binary crossover operator is as follows: for parent individuals and , the formula for generating offspring individuals and is as follows: ; ; wherein is a crossover parameter determined by a random number generator. The mutation operation adopts a polynomial mutation operator to randomly disturb the gene bits of the individuals.

[0074] The calculation formula of the mutation operator is as follows: ; wherein is the gene value after mutation, wherein denotes the gene value before mutation, and denote the upper and lower bounds of the gene, respectively, denotes the mutation strength, denotes a random number subject to a standard normal distribution.

[0075] Subsequently, the genetic algorithm is executed in cycles including the four stages of evaluation, selection, crossover, and mutation to perform iterations. The quality of each individual in the population is determined by fitness evaluation, the selection operation generates a mating pool, and the crossover and mutation operations generate new offspring individuals. The environmental selection mechanism selects excellent individuals from the parent and offspring individuals to form the next generation population, maintains the population size unchanged, and the termination condition is set as reaching the maximum number of iterations or the population convergence degree meeting the preset threshold. The final output candidate solution set contains non-dominated solutions on the Pareto frontier, and each solution represents a supply chain configuration scheme.

[0076] S43: The candidate solution set is subjected to feasibility verification by constraint condition filtering to obtain the dynamic candidate set.

[0077] In step S43, the present application performs feasibility verification on the candidate solution set through constraint filtering to obtain a dynamic candidate set, so as to ensure that the optimization result meets the actual business constraints. Constraint filtering is a constraint processing mechanism in the solution of a multi-objective optimization problem, which filters the feasible solutions by checking whether the candidate solutions meet the preset constraint conditions. The constraint conditions include equality constraints and inequality constraints. The equality constraints require that certain variables must satisfy an exact mathematical relationship, and the inequality constraints require that the values of certain variables must be within a specified range. The constraint conditions in the intelligent prediction and recommendation of a supply chain mainly include resource constraints, time constraints, capacity constraints, quality constraints, etc. The resource constraints require that the usage of various resources (such as raw materials, manpower, equipment, etc.) in the supply chain cannot exceed the available amount, the time constraints require that the execution time of the supply chain activities must meet the timing relationship, the capacity constraints require that the processing capacity of each link in the supply chain cannot exceed the maximum capacity, and the quality constraints require that the quality level of the supply chain products or services must meet the minimum standard.

[0078] The data processing process of constraint filtering includes constraint function calculation, feasibility judgment, and penalty function processing steps. The constraint function calculation process calculates the constraint violation degree of each candidate solution according to the mathematical expression of the constraint condition, which is used to measure the degree of deviation of the candidate solution from the constraint condition. The feasibility judgment process judges whether the candidate solution meets all the constraint conditions according to the constraint violation degree. Only the candidate solution that meets all the constraint conditions can enter the dynamic candidate set. The penalty function processing process punishes the candidate solution that violates the constraint condition, and reduces the fitness value of the candidate solution that violates the constraint by adding a penalty term to the objective function. The dynamic candidate set is a set of feasible solutions after constraint filtering. Each candidate solution in the set meets all the constraint conditions and has a good objective function value. The generation process of the dynamic candidate set also considers the dynamic changes of the supply chain environment, and adapts to the environmental changes by updating the constraint conditions and the objective function parameters in real time.

[0079] The maintenance process of the dynamic candidate set includes candidate solution evaluation, set updating, and diversity preservation steps. The candidate solution evaluation process comprehensively evaluates the candidate solutions according to the evaluation indexes of multi-objective optimization, which commonly include hyper-volume index, distance index, and distribution index, etc. The set updating process dynamically adjusts the candidate set according to the evaluation results, deletes the candidate solutions with poor quality, and adds new candidate solutions with good quality. The diversity preservation process maintains the diversity of the candidate set through diversity measurement and clustering analysis, avoiding the excessive aggregation of candidate solutions in a local optimal region. The size of the dynamic candidate set is usually controlled within a reasonable range, which ensures the quality and diversity of the candidate solutions, and avoids high computational complexity. The output format of the candidate set includes the decision variable values, objective function values, and constraint violation information of the candidate solutions, which provides input data for the subsequent personalized matching.

[0080] In a specific embodiment, the generation process of the dynamic candidate set of the present application is specifically illustrated by the electronic manufacturing industry supply chain recommendation scenario. An electronic product manufacturing enterprise needs to build a supply chain network for its new product line, including chip suppliers, electronic component suppliers, assembly factories, logistics service providers, and other links. The graph embedding vector contains a 256-dimensional feature representation of 2000 supply chain entities, and the numerical value of each vector element ranges between [-1, 1]. In step S41, the efficiency objective function is calculated based on a 12-month time window, the market demand data is derived from historical sales records and market prediction models, and the supply capacity data is based on supplier capacity evaluation and inventory statistics. The demand fluctuation decay coefficient is set to 0.1, indicating that the influence weight of each month is attenuated by about 10% relative to the previous month. The satisfaction objective function considers the demand satisfaction of 500 important customers, and the actual demand satisfaction rate is calculated by order completion rate and customer satisfaction survey data. The cost objective function includes 8 main cost items such as raw material procurement cost, production cost, logistics cost, and quality control cost, and the cost coefficient is determined based on historical cost data and supplier quotes. In step S42, the genetic algorithm sets the population size to 100, the maximum number of iterations to 200, the crossover probability to 0.8, and the mutation probability to 0.1. The individual encoding includes supplier selection, production allocation, logistics path, inventory strategy, and other decision variables, with a total dimension of 50. The initial population is generated by the Latin hypercube sampling method to ensure uniform distribution of the initial solution within the feasible region. After 200 generations of evolution calculation, the algorithm converges to a Pareto front containing 25 non-dominated solutions, each representing a supply chain configuration scheme that balances efficiency, satisfaction, and cost. In step S43, the constraint condition filtering process verifies 12 constraint conditions such as production capacity constraints, delivery time constraints, quality certification constraints, and fund budget constraints. Among them, 3 candidate solutions are excluded due to exceeding the production capacity limit, and 2 candidate solutions are filtered out due to not meeting the delivery time requirement, finally forming a dynamic candidate set containing 20 feasible solutions, providing high-quality candidate schemes that have undergone multi-objective optimization and constraint verification for subsequent personalized recommendation processing.

[0081] S5: performing personalized matching of user demand based on the dynamic candidate set to obtain a recommendation result.

[0082] Further, the dynamic candidate set is a candidate item set formed after the multi-objective optimization algorithm in the aforementioned step S4, which contains potential recommendation objects evaluated comprehensively in efficiency, satisfaction, and cost dimensions. These candidate items are represented as a structured data set containing supply chain entity identification, association relationship weight, time window attribute, and optimization score in data structure. Step S5 aims to convert the dynamic candidate set generated in the early stage into the final recommendation result.

[0083] The step S5 further comprises: S51: based on the dynamic candidate set, screening the supply chain layer candidate to obtain a supply chain layer recommendation set; filtering the merchant layer candidate to obtain a merchant layer recommendation set; and performing personalized screening on the user layer candidate to obtain a user layer recommendation set.

[0084] Further, the supply chain layer candidate screening process first needs to identify and classify the supply chain entities in the dynamic candidate set. The supply chain layer entities mainly include raw material suppliers, manufacturers, distributors, logistics service providers, and other upstream supply chain participants. The screening algorithm adopts a similarity matching mechanism based on a supply chain state feature vector, and determines the screening threshold by calculating the structural similarity between the candidate and the current supply chain network node. The generation of the supply chain layer recommendation set is specifically to construct an evaluation matrix to comprehensively evaluate the supply capacity, response speed, and stability indicators of the candidate.

[0085] The merchant layer candidate filtering process is a fine processing for the intermediate level of merchant entities. The merchant layer entities include retailers, wholesalers, e-commerce platforms, offline stores, and other direct market-oriented business entities. The filtering algorithm constructs multi-dimensional filtering conditions based on the operating scale, geographical location, customer group, product category, and other attribute features of the merchant. The merchant layer recommendation set in the present application comprehensively considers the transaction history, credit rating, inventory status, and distribution capacity of the merchant in multiple dimensions.

[0086] The user layer entity represents the final consumer or demand side, including individual users, enterprise users, institutional users, and other different types of demand subjects. The personalized screening algorithm needs to analyze the historical behavior data, preference features, consumption habits, geographical location, time preference, and other multi-dimensional information of the user. In the evaluation, the present application establishes a user feature vector to represent the attribute values of the user in multiple feature dimensions, wherein each dimension of the feature value is obtained by statistical analysis of the user historical interaction data and machine learning model prediction. Subsequently, the cosine similarity between the user feature vector and the candidate feature vector is calculated to determine the matching degree of the candidate and the user demand, so as to perform personalized screening.

[0087] S52: inputting the supply chain layer recommendation set, the merchant layer recommendation set, and the user layer recommendation set into a deep learning recommendation model to perform multi-task prediction and obtain a prediction score matrix.

[0088] Step S52 inputs the three-level recommendation set into the deep learning recommendation model for multi-task prediction processing. The deep learning recommendation model first receives the feature data of the supply chain layer recommendation set, the merchant layer recommendation set, and the user layer recommendation set through the input layer. The data format of each recommendation set is structured data containing entity identification, feature vector, and weight coefficient. The subsequent embedding layer converts different types of feature data into a unified dimension dense vector representation. The three levels of recommendation set are input into the deep learning recommendation model for multi-task prediction processing through the embedding matrix. The multi-task prediction mechanism simultaneously predicts different target tasks by designing multiple prediction heads. The prediction tasks include user satisfaction prediction, supply chain efficiency prediction, cost control prediction, and multiple dimensions. Each prediction task corresponds to an independent output layer, and information sharing between multiple tasks is realized through shared bottom layer feature representation. Finally, a prediction score matrix is obtained to represent the score results of a recommendation candidate in b prediction tasks.

[0089] S53: According to the prediction score matrix, the recommendation candidate is sorted and output by the sorting algorithm to obtain the recommendation result.

[0090] Further, step S53 sorts and outputs the recommendation candidate according to the prediction score matrix through the sorting algorithm. The sorting algorithm adopts a multi-dimensional score weighted sorting mechanism. First, the different dimension scores in the prediction score matrix need to be normalized. Then, the normalized score matrix is calculated by weighted summation to obtain the comprehensive score. Finally, the sorting algorithm sorts the candidate items in descending order according to the comprehensive score to generate a recommendation result list.

[0091] In one specific embodiment, the data processing process of the supply chain recommendation of the present application is illustrated by a specific example of automobile parts to explain its actual application. A certain automobile manufacturing enterprise needs to find suitable supply chain partners for its new vehicle model, including parts suppliers, logistics service providers, and dealer networks. The dynamic candidate set contains 1000 potential partners selected through the previous multi-objective optimization, including 300 parts suppliers, 200 logistics service providers, and 500 dealer candidates. In step S51, the supply chain layer filtering processes the 300 parts suppliers. The filtering conditions include production capacity, quality certification, on-time delivery rate, technical level, and other dimensions. The filtering algorithm calculates the comprehensive evaluation score of each supplier, and reduces the number of candidate suppliers to 80 high-quality suppliers that meet the requirements by setting the evaluation threshold.

[0092] The merchant layer filtering process conducts comprehensive evaluation on the 500 dealer candidates in terms of regional distribution, sales capacity, brand image, customer service level, etc. The filtering algorithm first filters the geographic location according to the market layout requirements of the automobile manufacturing enterprise, and then conducts comprehensive scoring based on the historical sales data, customer satisfaction evaluation, after-sales service quality, etc. of the dealers. After multiple rounds of filtering processing, 120 dealers with high-quality service capability and reasonable geographic distribution are finally selected to form the merchant layer recommendation set. The user layer personalized screening process differentiates the needs of different types of terminal consumers, including preference feature analysis of different user groups such as individual consumers, enterprise procurement users, and leasing companies.

[0093] In step S52, the deep learning recommendation model receives the recommendation set data containing 80 suppliers, 120 dealers, and 200 logistics service providers. The embedding layer of the model converts the multi-dimensional feature vector of each entity into a 128-dimensional dense representation. The hidden layer adopts a three-layer fully connected network structure, with 256 neurons in each layer. The multi-task prediction mechanism simultaneously predicts three key indicators: supply chain stability, cost-effectiveness, and service quality. After forward propagation calculation by the neural network, the model outputs the score results of the 400 candidates in the three prediction tasks, forming a 400x3 prediction score matrix.

[0094] The sorting process of step S53 first normalizes the prediction score matrix, unifying the scores in the three dimensions to the [0, 1] interval. Then, according to the business priority of the automobile manufacturing enterprise, the weight coefficients are set: the supply chain stability weight is 0.5, the cost-effectiveness weight is 0.3, and the service quality weight is 0.2. The weighted sum calculation obtains the comprehensive score of each candidate, and the sorting algorithm arranges the candidates in descending order according to the comprehensive score. At the same time, the diversity constraint is introduced to ensure that the recommendation results cover different types of partners. Finally, a recommendation result list containing the top 50 candidates is generated, providing a scientific basis for the selection of supply chain partners for automobile manufacturing enterprises.

[0095] As shown in Figure 2 The present application also provides an intelligent prediction and recommendation system based on supply chain full-link data, comprising: The acquisition module is used for acquiring multi-source heterogeneous supply chain data and performing standardization processing to obtain a standardized data set; The quantification module is used for quantitatively modeling the supply chain state characteristics according to the standardized data set to obtain a supply chain state feature vector; The modeling module is used for modeling the supply chain entity relationship based on the supply chain state feature vector through a heterogeneous graph neural network to obtain a graph embedding vector; The generation module is used for dynamically generating a candidate set through a multi-objective optimization algorithm according to the graph embedding vector to obtain a dynamic candidate set; The recommendation module is configured to perform personalized matching of user demand based on the dynamic candidate set to obtain a recommendation result.

[0096] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0097] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be realized by means of software and the necessary general hardware platform, and of course, it can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0098] The intelligent prediction recommendation method based on supply chain full-link data is constructed, and the intelligent upgrading of supply chain management is realized. First, the multi-source heterogeneous data fusion technology is adopted, the traditional supply chain data island problem is solved, and the data utilization efficiency is improved. Second, based on the entity relationship modeling of heterogeneous graph neural network, the complex node relationship and semantic information in the supply chain network are accurately captured, and the accuracy of the recommendation result is enhanced. Third, the application of multi-objective optimization algorithm realizes the comprehensive balance of efficiency, satisfaction and cost in three dimensions, and avoids the limitation of single objective optimization. In addition, the hierarchical recommendation architecture is differentially processed according to the demand characteristics of different levels of supply chain, and the pertinence and practicality of recommendation are improved. Finally, the multi-task prediction mechanism of deep learning model considers multiple business targets at the same time, and improves the scientificity and comprehensiveness of decision-making. Overall, the intelligent level of supply chain decision-making is significantly improved, the operation cost is reduced, and the supply chain response speed and customer satisfaction are improved.

[0099] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; 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.

Claims

1. An intelligent prediction recommendation method based on supply chain full-link data, characterized in that, include: S1: Collect multi-source heterogeneous supply chain data and perform standardization processing to obtain a standardized dataset; S2: Based on the standardized dataset, quantitatively model the supply chain status characteristics to obtain the supply chain status feature vector; S3: Based on the supply chain state feature vector, model the supply chain entity relationship through a heterogeneous graph neural network to obtain a graph embedding vector; S4: Based on the graph embedding vector, the candidate set is dynamically generated using a multi-objective optimization algorithm to obtain a dynamic candidate set; S5: Based on the dynamic candidate set, perform personalized matching of user needs to obtain recommendation results. 2.The intelligent prediction recommendation method based on supply chain whole-link data according to claim 1, characterized in that, Step S1 further includes: S11: Collect supply chain data to obtain raw data; the raw supply chain data includes: supplier raw data, logistics raw data, merchant raw data, and user raw data; S12: Deduplicate the original data to obtain cleaned data; S13: Convert the format of the cleaned data according to the unified data dictionary to obtain a standardized dataset. 3.The intelligent prediction recommendation method based on supply chain whole-link data according to claim 1, characterized in that, Step S2 further includes: S21: From the standardized dataset, the supply capacity dimension is extracted using a feature engineering framework to obtain supply capacity characteristics; from the standardized dataset, the demand fluctuation dimension is identified using time series analysis to obtain demand fluctuation characteristics; from the standardized dataset, the market trend dimension is mined using market analysis algorithms to obtain market trend characteristics. S22: The supply capacity characteristics, demand fluctuation characteristics, and market trend characteristics are weighted using a random forest algorithm, and the features are fused according to the assigned feature weight matrix to obtain fused features; S23: Using the fuzzy comprehensive evaluation method, the health of the supply chain is quantitatively assessed based on the fusion characteristics to obtain the supply chain status feature vector.

4. The intelligent prediction recommendation method based on supply chain full-link data according to claim 1, characterized in that, Step S3 further includes: S31: Model the nodes of the supply chain entities to obtain a supply chain heterogeneous graph; wherein, the nodes of the supply chain heterogeneous graph represent supply chain entities, and the edges of the supply chain heterogeneous graph represent supply chain relationships; S32: Perform semantic subgraph decomposition on the heterogeneous supply chain graph to obtain a set of semantic subgraphs; S33: Using a multilayer perceptron, the node features in the semantic subgraph set are projected to obtain a unified latent space representation; S34: Based on the unified latent space representation, multi-head attention of the meta-path weight matrix is ​​introduced to aggregate homogeneous neighbors and heterogeneous neighbors respectively to obtain neighbor aggregation features; S35: The neighbor aggregation features are fused through node-level attention and semantic-level attention to obtain the graph embedding vector.

5. The intelligent prediction recommendation method based on supply chain full-link data according to claim 4, characterized in that, Step S34 further includes: S341: Based on the aforementioned supply chain heterogeneous graph, discover relationship paths and obtain a meta-path set; S342: Perform path representation learning on the meta-path set to obtain meta-path embedding vectors; S343: Through similarity calculation, the importance of multiple meta-paths is evaluated based on the meta-path embedding vector to obtain the meta-path weight matrix; S344: Integrate the metapath weight matrix into multi-head attention to obtain fusion attention, aggregate the isomorphic neighbors and the heterogeneous neighbors through the fusion attention, and obtain neighbor aggregation features.

6. The intelligent prediction recommendation method based on supply chain full-link data according to claim 5, characterized in that, The expression of the fusion attention in step S344 is: ； wherein, is a meta-path index, is an other-path index, is a set of meta-paths, is a multi-head attention index, is a weight coefficient of a meta-path under a first head attention, is a query vector of a meta-path in a first head attention mechanism, is a key vector of a meta-path in a first head attention mechanism, is a sequence length, is a position encoding of a meta-path under a first head attention, is a query vector of an other-path in a first head attention mechanism, is a key vector of an other-path in a first head attention mechanism, is a position encoding of an other-path under a first head attention.

7. The intelligent prediction recommendation method based on supply chain full-link data according to claim 1, characterized in that, Step S4 further includes: S41: Construct an optimization objective function based on the graph embedding vector; S42: Perform global search on the optimization objective function through a genetic algorithm to obtain a candidate solution set; S43: Perform feasibility verification on the candidate solution set through constraint condition filtering to obtain the dynamic candidate set. 8.The intelligent prediction recommendation method based on supply chain whole-link data according to claim 7, characterized in that, The expression of the optimization objective function in step S41 is: ； wherein, is an efficiency objective function, is a weight coefficient corresponding to the efficiency objective function, is a satisfaction objective function, is a weight coefficient corresponding to the satisfaction objective function, is a cost objective function, is a weight coefficient corresponding to the cost objective function; ； wherein, denotes a time period, is the total number of time periods, is the market demand quantity of the time period, is the supply capacity of the time period, is the demand fluctuation decay coefficient, is the start of the current time window; ； wherein, is a user index value, is a total number of users, is an actual demand satisfaction rate of the th user, is a potential demand of the th user, is a user preference weight of the th user, denotes a time point, is a user behavior time point of the th user, is a time decay standard deviation; ； wherein, is an index value for a cost item, is a total number of cost items, is a cost coefficient of the th cost item, is a decision variable of the th cost item, is an uncertainty penalty factor, is a penalty coefficient for deviating from a reference value, is a reference value obtained by fitting historical data. 9.The intelligent prediction recommendation method based on supply chain whole-link data according to claim 1, characterized in that, Step S5 further includes: S51: Based on the dynamic candidate set, filter the supply chain layer candidate items to obtain a supply chain layer recommendation set, filter the merchant layer candidate items to obtain a merchant layer recommendation set, and perform personalized screening on the user layer candidate items to obtain a user layer recommendation set; S52: Input the supply chain layer recommendation set, the merchant layer recommendation set, and the user layer recommendation set into a deep learning recommendation model to perform multi-task prediction and obtain a prediction score matrix; S53: Sort the recommended candidate items according to the prediction score matrix through a sorting algorithm to output a recommendation result.

10. An intelligent prediction recommendation system based on supply chain full-link data, characterized in that, Comprise: The acquisition module is used for acquiring multi-source heterogeneous supply chain data and performing standardized processing to obtain a standardized data set; The quantification module is used for quantitatively modeling the supply chain state features according to the standardized data set to obtain a supply chain state feature vector; The modeling module is used for modeling the supply chain entity relationship based on the supply chain state feature vector through a heterogeneous graph neural network to obtain a graph embedding vector; The generation module is used for dynamically generating a candidate set through a multi-objective optimization algorithm according to the graph embedding vector to obtain a dynamic candidate set; The recommendation module is used for personalized matching of user demand based on the dynamic candidate set to obtain a recommendation result.

Images