Supply chain multidimensional data mining and intelligent recommendation decision-making method based on knowledge graph

Abstract

The present invention provides a supply chain multidimensional data mining and intelligent recommendation decision method based on a knowledge graph, which relates to the field of supply chain management technology, including constructing a dynamic supply chain knowledge graph containing the temporal evolution relationship of nodes, and calculating the multidimensional feature representation of nodes based on the graph, and training a supply chain recommendation model. Then, according to the target product parameters and constraints, the graph reasoning mechanism is used to calculate the correlation between the target product and the supplier, and the candidate suppliers are screened. Finally, combined with supplier evaluation and historical transaction data, a dynamic constraint optimization model is used to perform multi-objective optimization, and a recommended decision plan containing supplier ranking and matching degree is generated. The present invention can effectively integrate multidimensional data, realize accurate supplier recommendation, and improve supply chain efficiency and decision-making intelligence.

Description

Technical Field

[0001] The present invention relates to supply chain management technology, and in particular to a supply chain multidimensional data mining and intelligent recommendation decision-making method based on knowledge graph. Background Art

[0002] As supply chain networks continue to expand and become more complex, traditional supplier recommendation methods are unable to effectively process massive amounts of heterogeneous data and are unable to accurately characterize the dynamic evolution of supply chain systems. Existing supplier evaluation models based on static features ignore the dynamic changes in supplier capabilities and market environment, resulting in a large deviation between recommendation results and actual demand.

[0003] Currently, supply chain recommendation systems mainly rely on historical transaction data and fixed evaluation indicators to screen suppliers, lacking in-depth analysis of the complex relationships between suppliers and products. At the same time, due to the lack of a unified knowledge representation framework, it is difficult to fully explore and utilize the rich semantic information contained in multi-source heterogeneous data, which reduces the accuracy and interpretability of recommendation decisions.

[0004] Supply chain constraints are dynamically changing. Traditional constraint processing methods based on fixed thresholds have poor adaptability and cannot respond to market changes in a timely manner. Existing methods often separate supplier recommendation and constraint optimization, lacking a systematic decision optimization framework. Therefore, there is an urgent need for a supplier recommendation decision method that can integrate knowledge graphs and deep learning technologies to achieve dynamic analysis of multidimensional data in the supply chain, intelligent reasoning, and constraint optimization. Summary of the invention

[0005] The embodiment of the present invention provides a supply chain multidimensional data mining and intelligent recommendation decision-making method based on knowledge graph, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A supply chain multidimensional data mining and intelligent recommendation decision-making method based on knowledge graph is provided, including:

[0008] Supplier evaluation data, product evaluation data and supply and demand transaction data are collected and processed for text segmentation to obtain a word vector sequence, the context association of the word vector sequence is calculated to obtain a semantic feature vector, an initial supply chain knowledge graph is constructed based on the semantic feature vector, the supplier evaluation data, product evaluation data and supply and demand transaction data are used as attribute values ​​of corresponding nodes respectively, the temporal association strength between nodes is calculated to construct a multi-level topological structure, adjacent nodes are iteratively updated based on the multi-level topological structure, and a dynamic supply chain knowledge graph containing the temporal evolution relationship of nodes is generated;

[0009] Classify the nodes in the dynamic supply chain knowledge graph by type, calculate the information transfer weight between nodes according to the node type, aggregate the node features based on the information transfer weight to obtain a node feature matrix, divide the node feature matrix into multiple time windows according to the time series, calculate the state change of the node in each time window to obtain a time series feature vector, align and fuse the time series feature vector with the node feature matrix to obtain a multidimensional feature representation, calculate the importance weight of each dimensional feature in the multidimensional feature representation, select and reorganize the features according to the importance weight to obtain a fused feature vector, and train the supply chain recommendation model based on the fused feature vector;

[0010] The target product parameters and constraint condition parameters of the target product node are extracted from the dynamic supply chain knowledge graph and input into the supply chain recommendation model. The multi-hop path reachability between the target product node and the supplier node is calculated based on the graph reasoning mechanism to obtain the node association score. A set of candidate supplier nodes is screened out according to the node association score. The supplier evaluation data and historical transaction data of the candidate supplier node set are input into the dynamic constraint optimization model. The constraint boundary is updated through online learning and multi-objective optimization is performed to obtain the supplier optimization score. Based on the supplier optimization score, a recommended decision plan including supplier ranking and matching degree is generated.

[0011] In an optional embodiment,

[0012] Collecting supplier evaluation data, product evaluation data and supply and demand transaction data and performing text segmentation processing to obtain a word vector sequence, calculating the context association of the word vector sequence to obtain a semantic feature vector, and constructing an initial supply chain knowledge graph based on the semantic feature vector includes:

[0013] Perform word segmentation preprocessing on supplier evaluation data, product evaluation data and supply and demand transaction data to obtain a word vector sequence, calculate the frequency of occurrence of words in the word vector sequence in a single document and the distribution frequency in all document sets, construct a word weight calculation model to obtain a word weight matrix, extract a feature word sequence based on the word weight matrix, use a conditional random field model to annotate the feature word sequence to obtain an annotated word sequence, construct the annotated word sequence as an initial word library, construct a word dependency tree based on the initial word library to extract a syntactic dependency matrix between words, and perform hierarchical clustering on the initial word library according to the syntactic dependency matrix to obtain a dynamic word library;

[0014] Constructing a word co-occurrence matrix for the words in the dynamic vocabulary, calculating a word semantic similarity matrix based on the word co-occurrence matrix, using a bidirectional long short-term memory network to process the word semantic similarity matrix to extract context features to obtain a semantic representation vector group, constructing a time-aware mechanism to calculate a temporal weight coefficient, weighting the vectors in the semantic representation vector group according to the temporal weight coefficient to obtain a weighted vector group, and performing temporal fusion on the weighted vector group to obtain a semantic feature vector;

[0015] The supplier node, product node and transaction node are used as graph network vertices, the semantic feature vector is mapped into a node feature matrix, the node feature matrix is ​​processed and calculated by a graph attention network to obtain a multi-dimensional association probability matrix between nodes, a probability transfer matrix is ​​generated based on the multi-dimensional association probability matrix, and the node features are updated under the guidance of the probability transfer matrix through a message passing mechanism to obtain a node hidden layer feature matrix;

[0016] The matching coefficient matrix between nodes is calculated based on the node hidden layer feature matrix, wherein the matching coefficient matrix includes the supply and demand matching coefficient between the supplier node and the product node, the similarity coefficient between the product nodes, and the correlation coefficient between the supplier node and the transaction node. The matching coefficient matrix is ​​used as the initial weight matrix of the graph edge, and the random walk method is used to perform multiple rounds of iterative optimization on the initial weight matrix to obtain the final edge weight matrix. The initial supply chain knowledge graph is constructed based on the final edge weight matrix.

[0017] In an optional embodiment,

[0018] The supplier evaluation data, product evaluation data and supply and demand transaction data are respectively used as the attribute values ​​of the corresponding nodes, the temporal correlation strength between the nodes is calculated to construct a multi-level topological structure, and the adjacent nodes are iteratively updated based on the multi-level topological structure to generate a dynamic supply chain knowledge graph containing the temporal evolution relationship of the nodes, including:

[0019] The supplier evaluation data, product evaluation data and supply and demand transaction data are sorted by time series and divided into a forward sequence and a backward sequence, the forward sequence is forward time-series encoded to obtain a forward feature sequence, the backward sequence is reverse time-series encoded to obtain a backward feature sequence, the forward feature sequence and the backward feature sequence are concatenated to obtain supplier node time-series features, product node time-series features and transaction node time-series features, and the supplier node time-series features, product node time-series features and transaction node time-series features are respectively used as node attribute representations of corresponding nodes;

[0020] Calculating the node attribute variance within the time window based on the node attribute representation, calculating the time sensitivity in combination with the time interval between nodes, and using the time sensitivity to weight the degree of association between nodes to obtain the temporal association strength;

[0021] According to the temporal association strength, the nodes are hierarchically constructed into a multi-level topological structure, the spatial association degree between the nodes in the same layer in the multi-level topological structure is calculated to obtain a local association score, the temporal association degree between the nodes across layers is calculated to obtain a long-range association score, and the local association score and the long-range association score are integrated to obtain a comprehensive association strength between the nodes;

[0022] The historical state information of the node is stored, and the relevant historical state information is retrieved based on the current state of the node to obtain the node memory feature, and the states of the adjacent nodes are weighted and aggregated according to the comprehensive correlation strength between the nodes to obtain the node neighbor feature, and the node memory feature and the node neighbor feature are iteratively updated to obtain a new node state, and the stored historical state information is updated;

[0023] The updated node status is used as the dynamic feature representation of the node, and the node connection relationship is constructed according to the comprehensive correlation strength between the nodes to generate a dynamic supply chain knowledge graph containing the time-series evolution relationship of the nodes.

[0024] In an optional embodiment,

[0025] Classifying the nodes in the dynamic supply chain knowledge graph, calculating the information transfer weight between nodes according to the node type, aggregating the node features based on the information transfer weight to obtain a node feature matrix, dividing the node feature matrix into multiple time windows according to the time series, and calculating the state change of the node in each time window to obtain the time series feature vector includes:

[0026] Obtain the global statistical features and local structural features of the nodes in the dynamic supply chain knowledge graph, construct a multidimensional semantic space including business function dimensions, resource occupation dimensions and timing characteristics dimensions based on the global statistical features and local structural features, map the nodes to the multidimensional semantic space to obtain an initial semantic vector, use a multilayer perceptron to fuse the global statistical features and local structural features to obtain a dimensional weight coefficient, multiply the dimensional weight coefficient by the corresponding dimensional feature of the initial semantic vector to obtain a weighted semantic vector, and normalize the weighted semantic vector to obtain a node type representation;

[0027] Constructing a query vector and a key vector based on the node type representation, calculating the forward similarity between the query vector and the key vector between the node pair to obtain a forward attention score, calculating the reverse similarity between the key vector and the query vector between the node pair to obtain a reverse attention score, concatenating the forward attention score and the reverse attention score and inputting them into a gating unit to obtain a gating coefficient, and using the gating coefficient to perform a weighted combination of the forward attention score and the reverse attention score to obtain an information transfer weight between nodes;

[0028] Construct a multi-layer feature aggregation network, perform weighted aggregation of the features of adjacent nodes in each layer of the network according to the information transmission weights between the nodes, and map the aggregated features to a new feature space through an inter-layer conversion matrix to obtain the node representation of the current layer, superimpose the node representations of each layer to obtain a node feature matrix, and perform sequence division of the node feature matrix according to a preset time window size to obtain a time series feature sequence;

[0029] For each time window in the temporal feature sequence, a short-term convolution kernel, a medium-term convolution kernel and a long-term convolution kernel are used to perform convolution operations on the node feature matrix in the window to obtain short-term feature sequences, medium-term feature sequences and long-term feature sequences representing patterns of different time scales; an attention query matrix is ​​constructed based on the short-term feature sequence, medium-term feature sequence and long-term feature sequence, the similarity between the feature sequence and the attention query matrix is ​​calculated to obtain feature weights, and the feature weights are used to perform weighted combination of feature sequences of different scales to obtain a temporal feature vector.

[0030] In an optional embodiment,

[0031] Aligning and fusing the time series feature vector with the node feature matrix to obtain a multidimensional feature representation, calculating the importance weight of each dimensional feature in the multidimensional feature representation, selecting and reorganizing the features according to the importance weight to obtain a fused feature vector, and training a supply chain recommendation model based on the fused feature vector includes:

[0032] A bidirectional feature mapping network is constructed, and the input time series feature vector and node feature matrix are respectively nonlinearly transformed by the bidirectional feature mapping network to obtain projection features, the projection features with the same time series position are constructed as positive sample pairs, and the projection features and randomly selected projection features with different time series positions are constructed as negative sample pairs, and the feature similarity of the positive sample pair is maximized and the feature similarity of the negative sample pair is minimized by optimizing the contrast learning objective function to obtain alignment features;

[0033] Performing a linear transformation on the multidimensional feature representation to obtain a query matrix, a key matrix, and a value matrix, calculating the dot product of the query matrix and the key matrix to obtain an attention score, normalizing the attention score and multiplying it with the value matrix to obtain a feature importance weight matrix, applying sparse constraints and orthogonal constraints to the feature importance weight matrix to obtain feature weights;

[0034] According to the feature weights, each dimension of the multidimensional feature representation is sorted and screened to obtain important features, and the important features and the multidimensional feature representation are reconstructed by residual connection to obtain a fused feature vector;

[0035] Constraints are extracted from the supply chain system and converted into constraint vectors through a constraint encoding network. The fused feature vector and the constraint vector are inner-producted to obtain a recommendation score. A recommendation loss term is constructed based on the recommendation score. A constraint violation penalty term is constructed based on the degree of constraint satisfaction. A smoothing regularization term is added to construct a joint loss function. The joint loss function is optimized and trained using the gradient descent method to obtain a supply chain recommendation model.

[0036] In an optional embodiment,

[0037] Target product parameters and constraint condition parameters of the target product node are extracted from the dynamic supply chain knowledge graph and input into the supply chain recommendation model. The multi-hop path reachability between the target product node and the supplier node is calculated based on the graph reasoning mechanism to obtain the node association score. The candidate supplier node set is screened out according to the node association score, including:

[0038] Extract target product parameters and constraint condition parameters of the target product node from the dynamic supply chain knowledge graph, map the target product parameters to distribution vectors in the product semantic feature space, sample and generate target product features from the distribution vectors, decompose the constraint condition parameters to obtain a constraint condition feature sequence, construct a constraint condition vector according to the constraint condition feature sequence, and align the target product features with the constraint condition vector for probability distribution to obtain a target feature vector;

[0039] Construct a graph reasoning mechanism, take the target product node in the dynamic supply chain knowledge graph as the starting point, perform group path sampling based on the predefined meta-path pattern, group the sampled paths according to the relationship type and the number of hops to obtain a path group set, calculate the semantic consistency of the nodes in each path group to obtain the path group weight, sort the path group set by importance according to the path group weight, and select the top N path groups to construct the core path set of the target product node;

[0040] For each path in the core path set, the conditional probability between adjacent nodes on the path is calculated to obtain a node transfer probability sequence, a Markov inference model is constructed based on the node transfer probability sequence, a Monte Carlo method is used to perform random walk sampling on the Markov inference model to obtain a multi-hop path sampling set, and the occurrence frequency of each path in the multi-hop path sampling set is calculated to obtain a multi-hop path probability between a target product node and a supplier node;

[0041] Perform semantic matching on the target feature vector and the context information of the nodes on each path in the core path set to obtain a node importance score, perform weighted summation on the paths in the core path set based on the node importance score to obtain a path importance weight, and perform probability weighted fusion on the path importance weight and the multi-hop path probability to obtain a node association score between the target product node and the supplier node;

[0042] The node association scores are distributed and estimated to obtain a score distribution. A screening threshold is calculated based on the score distribution. The supplier nodes whose node association scores are greater than the screening threshold are taken as a preliminary set. Constraints are verified for the supplier nodes in the preliminary set, and the supplier nodes that pass the verification constitute a candidate supplier node set.

[0043] In an optional embodiment,

[0044] The supplier evaluation data and historical transaction data of the candidate supplier node set are input into the dynamic constraint optimization model, the constraint boundary is updated by online learning, and a multi-objective optimization is performed to obtain the supplier optimization score. Based on the supplier optimization score, a recommended decision plan including supplier ranking and matching degree is generated, including:

[0045] Extract supplier evaluation data and historical transaction data of a candidate supplier node set from a dynamic supply chain knowledge graph, construct a temporal cascade attention network to perform multi-scale feature extraction on the quality score, delivery score, and service score in the supplier evaluation data to obtain a comprehensive supplier score, use wavelet transform to perform multi-resolution decomposition on the historical transaction data to obtain trend features and cycle features, construct a transaction fluctuation matrix, a development trend matrix, and a transaction law tensor based on the trend features and cycle features, and obtain a transaction feature vector through tensor decomposition;

[0046] A hierarchical constraint graph structure is constructed according to the transaction feature vector, supply capacity constraint nodes, price range constraint nodes and delivery cycle constraint nodes are established in the hierarchical constraint graph structure, constraint violation data of multiple time windows are collected to construct a time-varying constraint violation graph, structural features of the time-varying constraint violation graph are extracted using a graph convolutional network to obtain a violation degree distribution, and a dynamic constraint boundary update mechanism is constructed based on the violation degree distribution to obtain a constraint condition update sequence;

[0047] Constructing a score maximization objective function based on the supplier comprehensive score, constructing a risk minimization objective function based on the transaction feature vector, and constructing a cost-benefit optimization objective function based on the constraint condition update sequence;

[0048] A multi-objective co-evolutionary optimization framework is constructed, the population is divided into sub-populations for parallel evolution, genes are exchanged between sub-populations through a competitive cooperation mechanism, and adaptive neighborhood search is used to enhance local search capabilities. A dynamic reference vector adjustment strategy is used to maintain population diversity. Multi-objective optimization is performed on the score maximization objective function, risk minimization objective function, and cost-effectiveness optimization objective function. The population is stratified and screened based on dominance level and crowding degree to obtain a multi-objective optimal solution set.

[0049] A multi-head interactive attention network is designed to calculate the dynamic weight of each objective function in the multi-objective optimal solution set, and the dynamic weight and the multi-objective optimal solution set are adaptively fused to obtain the supplier optimization score. The candidate supplier node set is sorted according to the supplier optimization score, and a fuzzy cognitive graph is constructed to calculate the satisfaction degree of each constraint condition and obtain the constraint satisfaction. The supplier optimization score and the constraint satisfaction are input into a deep coupling network to learn the mapping relationship between supplier matching. Based on the sorting results of the supplier optimization score and the supplier matching, a recommended decision plan including supplier sorting and matching is generated.

[0050] According to a second aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] According to a third aspect of the embodiments of the present invention,

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0057] In this embodiment, by analyzing text semantics and temporal associations, a dynamic supply chain knowledge graph containing node temporal evolution relationships is constructed, which can more accurately reflect the real situation and dynamic changes of the supply chain. Based on multidimensional feature representation and graph reasoning mechanism, it can more comprehensively consider the characteristics of suppliers and products, and combine historical transaction data and dynamic constraint optimization models to more accurately match supply and demand and improve the effectiveness of recommendations. It can automatically calculate node association scores and supplier optimization scores, and generate recommended decision plans including supplier ranking and matching, helping enterprises make intelligent decisions and optimize supplier selection and resource allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 The present invention is a flowchart of a method for multidimensional data mining and intelligent recommendation decision-making in a supply chain based on a knowledge graph. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1 FIG. 1 is a flow chart of a method for multidimensional data mining and intelligent recommendation decision-making in a supply chain based on a knowledge graph according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] S101. Collect supplier evaluation data, product evaluation data and supply and demand transaction data and perform text segmentation processing to obtain a word vector sequence, calculate the context association of the word vector sequence to obtain a semantic feature vector, construct an initial supply chain knowledge graph based on the semantic feature vector, use the supplier evaluation data, product evaluation data and supply and demand transaction data as the attribute values ​​of the corresponding nodes, calculate the temporal association strength between the nodes to construct a multi-level topological structure, iteratively update adjacent nodes based on the multi-level topological structure, and generate a dynamic supply chain knowledge graph containing the temporal evolution relationship of the nodes;

[0063] S102. Classify the nodes in the dynamic supply chain knowledge graph by type, calculate the information transfer weight between nodes according to the node type, aggregate the node features based on the information transfer weight to obtain a node feature matrix, divide the node feature matrix into multiple time windows according to the time series, calculate the state change of the node in each time window to obtain a time series feature vector, align and fuse the time series feature vector with the node feature matrix to obtain a multidimensional feature representation, calculate the importance weight of each dimensional feature in the multidimensional feature representation, select and reorganize the features according to the importance weight to obtain a fused feature vector, and train the supply chain recommendation model based on the fused feature vector;

[0064] S103. Target product parameters and constraint condition parameters of the target product node are extracted from the dynamic supply chain knowledge graph, and input into the supply chain recommendation model; the multi-hop path reachability between the target product node and the supplier node is calculated based on the graph reasoning mechanism to obtain the node association score; a set of candidate supplier nodes is screened out according to the node association score; the supplier evaluation data and historical transaction data of the candidate supplier node set are input into the dynamic constraint optimization model; the constraint boundary is updated through online learning and multi-objective optimization is performed to obtain the supplier optimization score; and a recommended decision plan including supplier ranking and matching degree is generated based on the supplier optimization score.

[0065] In an optional implementation, supplier evaluation data, product evaluation data, and supply and demand transaction data are collected and text segmentation is performed to obtain a word vector sequence, context association of the word vector sequence is calculated to obtain a semantic feature vector, and an initial supply chain knowledge graph is constructed based on the semantic feature vector, including:

[0066] Perform word segmentation preprocessing on supplier evaluation data, product evaluation data and supply and demand transaction data to obtain a word vector sequence, calculate the frequency of occurrence of words in the word vector sequence in a single document and the distribution frequency in all document sets, construct a word weight calculation model to obtain a word weight matrix, extract a feature word sequence based on the word weight matrix, use a conditional random field model to annotate the feature word sequence to obtain an annotated word sequence, construct the annotated word sequence as an initial word library, construct a word dependency tree based on the initial word library to extract a syntactic dependency matrix between words, and perform hierarchical clustering on the initial word library according to the syntactic dependency matrix to obtain a dynamic word library;

[0067] Constructing a word co-occurrence matrix for the words in the dynamic vocabulary, calculating a word semantic similarity matrix based on the word co-occurrence matrix, using a bidirectional long short-term memory network to process the word semantic similarity matrix to extract context features to obtain a semantic representation vector group, constructing a time-aware mechanism to calculate a temporal weight coefficient, weighting the vectors in the semantic representation vector group according to the temporal weight coefficient to obtain a weighted vector group, and performing temporal fusion on the weighted vector group to obtain a semantic feature vector;

[0068] The supplier node, product node and transaction node are used as graph network vertices, the semantic feature vector is mapped into a node feature matrix, the node feature matrix is ​​processed and calculated by a graph attention network to obtain a multi-dimensional association probability matrix between nodes, a probability transfer matrix is ​​generated based on the multi-dimensional association probability matrix, and the node features are updated under the guidance of the probability transfer matrix through a message passing mechanism to obtain a node hidden layer feature matrix;

[0069] The matching coefficient matrix between nodes is calculated based on the node hidden layer feature matrix, wherein the matching coefficient matrix includes the supply and demand matching coefficient between the supplier node and the product node, the similarity coefficient between the product nodes, and the correlation coefficient between the supplier node and the transaction node. The matching coefficient matrix is ​​used as the initial weight matrix of the graph edge, and the random walk method is used to perform multiple rounds of iterative optimization on the initial weight matrix to obtain the final edge weight matrix. The initial supply chain knowledge graph is constructed based on the final edge weight matrix.

[0070] Exemplarily, firstly, supplier evaluation data, product evaluation data and supply and demand transaction data are collected, for example, from e-commerce platforms, internal enterprise databases, etc. These data can be text-based evaluations, transaction records, etc.

[0071] The collected data is pre-processed by word segmentation. Taking the supplier evaluation data as an example, if there is a review that says "this supplier has a very fast delivery speed and good product quality", after word segmentation, we get "this / supplier / delivery / speed / very / fast / product / quality / is / very / good". Similar word segmentation is also performed on product evaluation data and supply and demand transaction data.

[0072] Calculate word weights. Count the frequency of each word in a single document and the distribution frequency in all document sets. For example, "supplier" appears once in this review. Assuming that it appears 1,000 times in all supplier review data and the total number of documents is 10,000, the weight of the word "supplier" can be calculated using algorithms such as TF-IDF. Perform similar weight calculations on other words and finally obtain a word weight matrix.

[0073] Extract feature word sequences based on word weight matrix. Set a weight threshold, filter out words with weight greater than the threshold, and form feature word sequences. For example, filter out the top 1000 words in weight ranking.

[0074] The conditional random field model is used to annotate the feature word sequences. For example, "supplier" is annotated as "SUPPLYER", "product" is annotated as "PRODUCT", and "delivery speed" is annotated as "DELIVERY_SPEED". All annotated word sequences are constructed as the initial word library.

[0075] Build a word dependency tree based on the initial word library. For example, for "This supplier delivers goods quickly", a dependency tree can be built to represent the syntactic relationship between words, such as "supplier" is the subject of "delivery speed". Extract the syntactic dependency matrix between words, which records the dependency relationship between words.

[0076] The initial word library is hierarchically clustered according to the syntactic dependency matrix. For example, words with similar dependencies are clustered together, such as "shipping speed" and "logistics speed". After clustering, a dynamic word library is obtained, which can be dynamically updated according to new data.

[0077] Construct a word co-occurrence matrix for the words in the dynamic word library. Count the number of times two words co-occur in the same document. For example, if "supplier" and "product" co-occur 100 times in the same document, the value of the corresponding position in the co-occurrence matrix is ​​100.

[0078] The word semantic similarity matrix is ​​calculated based on the word co-occurrence matrix. For example, the semantic similarity between two words can be calculated using methods such as cosine similarity.

[0079] The bidirectional long short-term memory network is used to process the word semantic similarity matrix. The semantic similarity matrix is ​​used as input to extract context features and obtain a semantic representation vector group. For example, each word corresponds to a semantic representation vector.

[0080] A time-aware mechanism is built to calculate the time series weight coefficients. For example, more weight is given to the most recent data and less weight is given to the older data.

[0081] The vectors in the semantic representation vector group are weighted according to the temporal weight coefficient. Each vector is multiplied by the corresponding temporal weight coefficient.

[0082] The weighted vector group is temporally fused to obtain a semantic feature vector. For example, the weighted vectors may be averaged or weighted averaged to obtain a final semantic feature vector.

[0083] Suppliers, products, and transactions are used as graph network vertices. For example, supplier A, product B, and transaction C are used as three vertices respectively.

[0084] Map the semantic feature vector to a node feature matrix. For example, the semantic feature vector of supplier A is used as the feature vector of the supplier A node.

[0085] The node feature matrix is ​​processed using a graph attention network. The multi-dimensional association probability matrix between nodes is calculated. This matrix represents the association probability between nodes in different dimensions.

[0086] A probability transfer matrix is ​​generated based on the multi-dimensional association probability matrix. This matrix is ​​used to guide message delivery.

[0087] Update node features through message passing mechanism. Under the guidance of probability transfer matrix, nodes pass information to each other, update node features, and obtain node hidden layer feature matrix.

[0088] The matching coefficient matrix between nodes is calculated based on the node hidden layer feature matrix. For example, the supply-demand matching coefficient between supplier A and product B, the similarity coefficient between product B and product C, and the correlation coefficient between supplier A and transaction C are calculated.

[0089] The matching coefficient matrix is ​​used as the initial weight matrix of the graph edges.

[0090] The random walk method is used to perform multiple rounds of iterative optimization on the initial weight matrix. After multiple iterations, the final edge weight matrix is ​​obtained.

[0091] Construct the initial supply chain knowledge graph based on the final edge weight matrix. Connect the nodes and edges to form a knowledge graph.

[0092] Suppose there is a supplier evaluation data "This supplier has fast delivery and good product quality", a product evaluation data "This product has excellent performance and reasonable price", and a supply and demand transaction data "Supplier A provides product C to enterprise B". After the above steps, we can get the supplier node "Supplier A", the product node "Product C", and the supply and demand relationship between them.

[0093] In this embodiment, the word library is dynamically constructed through word segmentation, weight calculation and hierarchical clustering, which can accurately capture domain-specific semantic features; combined with conditional random fields and bidirectional long short-term memory networks, contextual information is fully extracted and the model's time perception ability is enhanced. The node feature learning and probability transfer mechanism based on the graph attention network not only strengthens the multi-dimensional association between suppliers, products and transaction nodes, but also adaptively optimizes supply and demand matching and correlation calculation. Finally, by iteratively optimizing edge weights through random walks, a more accurate and dynamic supply chain knowledge graph is formed, thereby improving the intelligence, prediction accuracy and collaborative efficiency of supply chain management.

[0094] In an optional implementation, the supplier evaluation data, product evaluation data and supply and demand transaction data are respectively used as attribute values ​​of corresponding nodes, the temporal correlation strength between nodes is calculated to construct a multi-level topological structure, and adjacent nodes are iteratively updated based on the multi-level topological structure to generate a dynamic supply chain knowledge graph containing the temporal evolution relationship of nodes, including:

[0095] The supplier evaluation data, product evaluation data and supply and demand transaction data are sorted by time series and divided into a forward sequence and a backward sequence, the forward sequence is forward time-series encoded to obtain a forward feature sequence, the backward sequence is reverse time-series encoded to obtain a backward feature sequence, the forward feature sequence and the backward feature sequence are concatenated to obtain supplier node time-series features, product node time-series features and transaction node time-series features, and the supplier node time-series features, product node time-series features and transaction node time-series features are respectively used as node attribute representations of corresponding nodes;

[0096] Calculating the node attribute variance within the time window based on the node attribute representation, calculating the time sensitivity in combination with the time interval between nodes, and using the time sensitivity to weight the degree of association between nodes to obtain the temporal association strength;

[0097] According to the temporal association strength, the nodes are hierarchically constructed into a multi-level topological structure, the spatial association degree between the nodes in the same layer in the multi-level topological structure is calculated to obtain a local association score, the temporal association degree between the nodes across layers is calculated to obtain a long-range association score, and the local association score and the long-range association score are integrated to obtain a comprehensive association strength between the nodes;

[0098] The historical state information of the node is stored, and the relevant historical state information is retrieved based on the current state of the node to obtain the node memory feature, and the states of the adjacent nodes are weighted and aggregated according to the comprehensive correlation strength between the nodes to obtain the node neighbor feature, and the node memory feature and the node neighbor feature are iteratively updated to obtain a new node state, and the stored historical state information is updated;

[0099] The updated node status is used as the dynamic feature representation of the node, and the node connection relationship is constructed according to the comprehensive correlation strength between the nodes to generate a dynamic supply chain knowledge graph containing the time-series evolution relationship of the nodes.

[0100] Exemplarily, first, the data is preprocessed. The supplier evaluation data, product evaluation data, and supply and demand transaction data are sorted in chronological order. In order to capture the bidirectional information of the time series, the sorted data is divided into a forward sequence and a backward sequence. For example, after sorting the monthly data from January 2022 to December 2023, the data from January 2022 to June 2023 can be used as the forward sequence, and the data from December 2023 to July 2023 can be used as the backward sequence. Afterwards, a temporal encoding method such as Word2Vec or BERT is used to forward encode the forward sequence to obtain a forward feature sequence; the backward sequence is reversely encoded to obtain a backward feature sequence. For example, Word2Vec can be used to encode the evaluation data of each month into a vector. Finally, the forward feature sequence and the backward feature sequence are spliced ​​together to obtain the temporal features of the supplier, product, and transaction nodes respectively, as the attribute representation of the node. Assuming that the forward feature sequence is [0.1, 0.2, 0.3] and the backward feature sequence is [0.4, 0.5, 0.6], the concatenated node attributes are expressed as [0.1, 0.2, 0.3, 0.4, 0.5, 0.6].

[0101] Next, calculate the temporal association strength between nodes. First, calculate the attribute variance of each node within a certain time window. For example, calculate the variance of the attribute representation vector of each supplier node in the past three months. Then, calculate the time sensitivity in combination with the time interval between nodes. The shorter the time interval, the higher the time sensitivity. For example, a decreasing function can be used to represent the relationship between time sensitivity and time interval. Finally, use time sensitivity to weight the degree of association between nodes to obtain the temporal association strength. For example, the degree of association between two supplier nodes can be measured by the cosine similarity of their attribute representation vectors, and then multiplied by the time sensitivity to obtain the final temporal association strength.

[0102] Then, a multi-level topological structure is constructed. The nodes are layered according to the calculated temporal correlation strength. For example, the nodes can be divided into a core layer, an intermediate layer, and an edge layer according to the magnitude of the temporal correlation strength. Then, the spatial correlation degree between nodes in the same layer is calculated, for example, the similarity of the attribute representation of nodes in the same layer is calculated to obtain a local correlation score. At the same time, the temporal correlation degree between nodes across layers is calculated, for example, the temporal correlation strength between core layer nodes and edge layer nodes is calculated to obtain a long-range correlation score. Finally, the local correlation score and the long-range correlation score are fused, for example, by weighted averaging, to obtain a comprehensive correlation strength between nodes.

[0103] After that, the node status is iteratively updated. First, the historical status information of each node is stored, such as the attribute representation and association strength in the past period of time. Then, based on the current state of the node, the relevant historical status information is retrieved to obtain the node memory feature. Next, the states of adjacent nodes are weighted and aggregated according to the comprehensive association strength between nodes to obtain the node neighbor feature. For example, the neighbor feature of a node can be represented by the weighted average of the states of its neighbor nodes, and the weight is the comprehensive association strength between nodes. Finally, the node memory feature and the node neighbor feature are iteratively updated to obtain a new node state and update the stored historical status information.

[0104] Finally, a dynamic supply chain knowledge graph is generated. The updated node status is used as the dynamic feature representation of the node. The node connection relationship is constructed according to the comprehensive association strength between nodes. For example, nodes whose comprehensive association strength exceeds a certain threshold can be connected. Finally, a dynamic supply chain knowledge graph containing the temporal evolution relationship of nodes is generated.

[0105] In this embodiment, by capturing the temporal evolution relationship of nodes, potential supply chain risks, such as supplier closures and product quality issues, can be predicted more accurately. Through dynamic knowledge graphs, the operation mechanism of the supply chain can be better understood, key nodes and paths can be identified, and resource allocation can be optimized to improve operational efficiency. Dynamic knowledge graphs can provide more comprehensive information and deeper insights for supply chain management, supporting smarter decisions, such as supplier selection and inventory management.

[0106] In an optional implementation, the nodes in the dynamic supply chain knowledge graph are classified by type, the information transfer weight between nodes is calculated according to the node type, the node features are aggregated based on the information transfer weight to obtain a node feature matrix, the node feature matrix is ​​divided into multiple time windows according to the time series, and the state change of the node in each time window is calculated to obtain a time series feature vector, including:

[0107] Obtain the global statistical features and local structural features of the nodes in the dynamic supply chain knowledge graph, construct a multidimensional semantic space including business function dimensions, resource occupation dimensions and timing characteristics dimensions based on the global statistical features and local structural features, map the nodes to the multidimensional semantic space to obtain an initial semantic vector, use a multilayer perceptron to fuse the global statistical features and local structural features to obtain a dimensional weight coefficient, multiply the dimensional weight coefficient by the corresponding dimensional feature of the initial semantic vector to obtain a weighted semantic vector, and normalize the weighted semantic vector to obtain a node type representation;

[0108] Constructing a query vector and a key vector based on the node type representation, calculating the forward similarity between the query vector and the key vector between the node pair to obtain a forward attention score, calculating the reverse similarity between the key vector and the query vector between the node pair to obtain a reverse attention score, concatenating the forward attention score and the reverse attention score and inputting them into a gating unit to obtain a gating coefficient, and using the gating coefficient to perform a weighted combination of the forward attention score and the reverse attention score to obtain an information transfer weight between nodes;

[0109] Construct a multi-layer feature aggregation network, perform weighted aggregation of the features of adjacent nodes in each layer of the network according to the information transmission weights between the nodes, and map the aggregated features to a new feature space through an inter-layer conversion matrix to obtain the node representation of the current layer, superimpose the node representations of each layer to obtain a node feature matrix, and perform sequence division of the node feature matrix according to a preset time window size to obtain a time series feature sequence;

[0110] For each time window in the temporal feature sequence, a short-term convolution kernel, a medium-term convolution kernel and a long-term convolution kernel are used to perform convolution operations on the node feature matrix in the window to obtain short-term feature sequences, medium-term feature sequences and long-term feature sequences representing patterns of different time scales; an attention query matrix is ​​constructed based on the short-term feature sequence, medium-term feature sequence and long-term feature sequence, the similarity between the feature sequence and the attention query matrix is ​​calculated to obtain feature weights, and the feature weights are used to perform weighted combination of feature sequences of different scales to obtain a temporal feature vector.

[0111] Exemplarily, first, the nodes in the dynamic supply chain knowledge graph are classified by type. The global statistical features of each node in the knowledge graph, such as the degree and centrality of the node, and the local structural features, such as the type distribution of the node's neighbors and the type of relationship in which the node participates, are obtained. Based on these features, a multidimensional semantic space containing business function dimensions, resource occupation dimensions, and timing characteristics dimensions is constructed. Each node is mapped to this multidimensional semantic space to obtain an initial semantic vector. Then, a multilayer perceptron is used to fuse the global statistical features and local structural features to learn the dimension weight coefficient. The dimension weight coefficient is multiplied by the corresponding dimension feature of the initial semantic vector to obtain a weighted semantic vector. Finally, the weighted semantic vector is normalized to obtain a node type representation. For example, the global statistical features of a node are height and high centrality, and the local structural features are mainly connected to the supplier node and participate in the "supply" relationship. After weighting and normalization, its vector in the multidimensional semantic space can be identified as a "core supplier" type.

[0112] Next, the information transfer weight between nodes is calculated based on the node type representation. Based on the node type representation, the query vector and key vector of each node are constructed. The forward similarity between the query vector and the key vector between the node pairs is calculated to obtain the forward attention score. For example, if the query vector of node A has a higher similarity with the key vector of node B, the forward attention score is higher. At the same time, the reverse similarity between the key vector and the query vector between the node pairs is calculated to obtain the reverse attention score. The forward attention score and the reverse attention score are concatenated and input into the gating unit to obtain the gating coefficient. The gating coefficient is used to perform a weighted combination of the forward attention score and the reverse attention score to obtain the information transfer weight between nodes. For example, if node A has a higher positive attention to node B, and node B has a lower reverse attention to node A, then after the gating mechanism, the final information transfer weight from A to B will be reduced.

[0113] Then, the node features are aggregated based on the information transfer weight to obtain a node feature matrix. A multi-layer feature aggregation network is constructed. In each layer of the network, the features of adjacent nodes are weighted and aggregated according to the information transfer weight between nodes, and the aggregated features are mapped to the new feature space through the inter-layer conversion matrix to obtain the node representation of the current layer. The node representations of each layer are superimposed to obtain a node feature matrix. For example, when the neighbor features of node A are aggregated in the first layer, a weighted sum is performed based on the information transfer weight between A and its neighbors. The result is then mapped to a new feature space through the conversion matrix. After repeating the multi-layer operation, the results of each layer are superimposed to obtain the final feature representation of node A. The feature representations of all nodes are combined to form a node feature matrix. The node feature matrix is ​​sequenced according to the preset time window size to obtain a time series feature sequence. For example, a node feature matrix of one month is divided into a time window per week to obtain a four-week time series feature sequence.

[0114] Finally, the state change of the node in each time window is calculated to obtain the time series feature vector. For each time window in the time series feature sequence, the short-term convolution kernel, the medium-term convolution kernel and the long-term convolution kernel are used to perform convolution operations on the node feature matrix in the window to obtain short-term feature sequences, medium-term feature sequences and long-term feature sequences that represent different time scale patterns. For example, a convolution kernel with a length of 1 is used to extract short-term features, a convolution kernel with a length of 3 is used to extract medium-term features, and a convolution kernel with a length of 7 is used to extract long-term features. An attention query matrix is ​​constructed based on the short-term feature sequence, the medium-term feature sequence and the long-term feature sequence. The similarity between the feature sequence and the attention query matrix is ​​calculated to obtain the feature weight. The feature weight is used to perform a weighted combination of feature sequences of different scales to obtain a time series feature vector. For example, if the short-term feature in a time window fluctuates greatly, the short-term feature weight will be higher, so that more attention is paid to the short-term change pattern.

[0115] In this embodiment, it is possible to achieve a fine characterization of the dynamic changes in the supply chain and extract multi-scale temporal features, significantly improving the accuracy of node feature representation and context awareness. By fusing global statistical and local structural features, supply chain nodes are mapped to a multidimensional semantic space to capture business, resource, and timing characteristics. The forward and reverse attention mechanisms and gating units effectively enhance the accuracy and stability of information transmission between nodes. The multi-layer feature aggregation network weights and aggregates node features layer by layer, and extracts short-term, medium-term, and long-term patterns through convolution kernels of different time scales. The attention mechanism is combined to achieve dynamic weighting of features, ensuring the model's sensitivity to the timing laws of the supply chain and the reliability of predictions.

[0116] In an optional implementation, the time series feature vector is aligned and fused with the node feature matrix to obtain a multidimensional feature representation, the importance weight of each dimensional feature in the multidimensional feature representation is calculated, the features are selected and reorganized according to the importance weight to obtain a fused feature vector, and the supply chain recommendation model is obtained by training based on the fused feature vector, including:

[0117] A bidirectional feature mapping network is constructed, and the input time series feature vector and node feature matrix are respectively nonlinearly transformed by the bidirectional feature mapping network to obtain projection features, the projection features with the same time series position are constructed as positive sample pairs, and the projection features and randomly selected projection features with different time series positions are constructed as negative sample pairs, and the feature similarity of the positive sample pair is maximized and the feature similarity of the negative sample pair is minimized by optimizing the contrast learning objective function to obtain alignment features;

[0118] Performing a linear transformation on the multidimensional feature representation to obtain a query matrix, a key matrix, and a value matrix, calculating the dot product of the query matrix and the key matrix to obtain an attention score, normalizing the attention score and multiplying it with the value matrix to obtain a feature importance weight matrix, applying sparse constraints and orthogonal constraints to the feature importance weight matrix to obtain feature weights;

[0119] According to the feature weights, each dimension of the multidimensional feature representation is sorted and screened to obtain important features, and the important features and the multidimensional feature representation are reconstructed by residual connection to obtain a fused feature vector;

[0120] Constraints are extracted from the supply chain system and converted into constraint vectors through a constraint encoding network. The fused feature vector and the constraint vector are inner-producted to obtain a recommendation score. A recommendation loss term is constructed based on the recommendation score. A constraint violation penalty term is constructed based on the degree of constraint satisfaction. A smoothing regularization term is added to construct a joint loss function. The joint loss function is optimized and trained using the gradient descent method to obtain a supply chain recommendation model.

[0121] The construction method of the supply chain recommendation model aims to provide a more accurate and efficient supply chain recommendation service. This method extracts key features by aligning and fusing time series features and node features, and trains the model in combination with supply chain constraints, ultimately achieving optimization of recommendation results.

[0122] For example, first, multi-source heterogeneous data such as historical transaction data, product information, and supplier information are collected from the supply chain system. For example, information such as the order volume, product category, supplier production capacity, and credit rating for each month in the past year is collected. The time series feature vector can be expressed as the change trend of the monthly order volume, and the node feature matrix can represent the attribute information of products and suppliers.

[0123] Then, a bidirectional feature mapping network is constructed, which consists of multiple fully connected layers. The time series feature vector and the node feature matrix are respectively input into the network for nonlinear transformation to obtain the corresponding projection features. Assuming that the dimension of the time series feature vector is 12 (representing 12 months) and the dimension of the node feature matrix is ​​50 (representing 50 suppliers), after nonlinear transformation, the dimension of the projection features is 64. The projection features of the same month are regarded as positive sample pairs. For example, the time series projection features of January and the node projection features of January constitute a positive sample pair. The random combination of projection features of different months is regarded as a negative sample pair. For example, the time series projection features of January and the node projection features of March constitute a negative sample pair. Through contrastive learning, the feature similarity of the positive sample pair is maximized, and the feature similarity of the negative sample pair is minimized, and finally the aligned features are obtained.

[0124] Next, the aligned features are concatenated with the original time series feature vectors and node feature matrices to obtain a multidimensional feature representation. This multidimensional feature representation is linearly transformed to obtain the query matrix, key matrix, and value matrix. The dot product of the query matrix and the key matrix is ​​calculated to obtain the attention score. The attention score is normalized and then multiplied with the value matrix to obtain the feature importance weight matrix. Sparse constraints and orthogonal constraints are imposed on the matrix to obtain the final feature weights. For example, assuming that the dimension of the multidimensional feature representation is 128, after linear transformation, the dimensions of the query matrix, key matrix, and value matrix are all 128. After a series of calculations, 128 feature weights are obtained, representing the importance of each feature.

[0125] According to the feature weights, the importance of each dimension of the multidimensional feature representation is sorted and screened to retain the important features. These important features are reconstructed with the original multidimensional feature representation through residual connection to obtain a fused feature vector. For example, the top 64 features with the weight ranking are selected as important features, and then these 64 features are connected with the original 128-dimensional feature vector through residual connection to obtain the final fused feature vector.

[0126] Extract constraints from the supply chain system, such as the capacity constraints of suppliers, product demand forecasts, etc. Convert these constraints into constraint vectors through the constraint encoding network. Perform inner product operation on the fused feature vector and the constraint vector to obtain the recommendation score. Construct the recommendation loss term based on the recommendation score, construct the constraint violation penalty term based on the degree of constraint satisfaction, and add a smoothing regularization term to construct a joint loss function. Use the gradient descent method to optimize the joint loss function and finally obtain the supply chain recommendation model.

[0127] In this embodiment, contrastive learning is used to maximize the temporal consistency of positive sample pairs and minimize the differences of negative sample pairs, thereby obtaining more robust alignment features. The attention mechanism is combined with sparse and orthogonal constraints to effectively screen out the most important features with business value, and achieve feature reconstruction through residual connections to enhance the generalization ability of the model. The constraint coding network integrates the hard and soft constraints in the supply chain system into the recommendation process, so that the model not only focuses on maximizing the recommendation score, but also takes into account the strict compliance of business rules. The joint loss function strikes a balance between recommendation accuracy, constraint satisfaction, and smooth regularization, and optimizes training through gradient descent to effectively avoid overfitting and improve the stability of the model, and finally constructs a recommendation result that can dynamically adapt to supply chain fluctuations and output high-quality, business-compliant results.

[0128] In an optional implementation, target product parameters and constraint condition parameters of the target product node are extracted from the dynamic supply chain knowledge graph and input into the supply chain recommendation model. The multi-hop path reachability between the target product node and the supplier node is calculated based on the graph reasoning mechanism to obtain a node association score. The candidate supplier node set is screened out according to the node association score, including:

[0129] Extract target product parameters and constraint condition parameters of the target product node from the dynamic supply chain knowledge graph, map the target product parameters to distribution vectors in the product semantic feature space, sample and generate target product features from the distribution vectors, decompose the constraint condition parameters to obtain a constraint condition feature sequence, construct a constraint condition vector according to the constraint condition feature sequence, and align the target product features with the constraint condition vector for probability distribution to obtain a target feature vector;

[0130] Construct a graph reasoning mechanism, take the target product node in the dynamic supply chain knowledge graph as the starting point, perform group path sampling based on the predefined meta-path pattern, group the sampled paths according to the relationship type and the number of hops to obtain a path group set, calculate the semantic consistency of the nodes in each path group to obtain the path group weight, sort the path group set by importance according to the path group weight, and select the top N path groups to construct the core path set of the target product node;

[0131] For each path in the core path set, the conditional probability between adjacent nodes on the path is calculated to obtain a node transfer probability sequence, a Markov inference model is constructed based on the node transfer probability sequence, a Monte Carlo method is used to perform random walk sampling on the Markov inference model to obtain a multi-hop path sampling set, and the occurrence frequency of each path in the multi-hop path sampling set is calculated to obtain a multi-hop path probability between a target product node and a supplier node;

[0132] Perform semantic matching on the target feature vector and the context information of the nodes on each path in the core path set to obtain a node importance score, perform weighted summation on the paths in the core path set based on the node importance score to obtain a path importance weight, and perform probability weighted fusion on the path importance weight and the multi-hop path probability to obtain a node association score between the target product node and the supplier node;

[0133] The node association scores are distributed and estimated to obtain a score distribution. A screening threshold is calculated based on the score distribution. The supplier nodes whose node association scores are greater than the screening threshold are taken as a preliminary set. Constraints are verified for the supplier nodes in the preliminary set, and the supplier nodes that pass the verification constitute a candidate supplier node set.

[0134] For example, first, extract the parameters and constraints of the target product from the dynamic supply chain knowledge graph. For example, if the target product is "electric vehicle battery", extract its parameters such as "battery capacity", "energy density", "cycle life", etc., as well as constraints such as "delivery cycle less than 30 days", "warranty period not less than 5 years", "suppliers must pass ISO9001 certification", etc.

[0135] Next, the target product parameters are mapped to the product semantic feature space and converted into a distributed vector representation. For example, "battery capacity" is mapped to a numerical interval to represent batteries of different capacity levels, and is converted into a vector using word embedding technology. Then, multiple target product features are sampled from the distribution vector. For example, feature vectors representing batteries of different capacity levels can be sampled. At the same time, the constraints are decomposed, for example, "delivery cycle is less than 30 days" is decomposed into two features, "delivery cycle" and "30 days", and converted into a constraint feature sequence. The constraint vector is constructed based on the sequence, for example, the feature vectors of "delivery cycle" and "30 days" are concatenated. Finally, the target product features are probability-aligned with the constraint vector, for example, the cosine similarity between them is calculated to obtain the final target feature vector.

[0136] Then, a graph reasoning mechanism is constructed. Starting from the target product node, group path sampling is performed based on predefined meta-path patterns (e.g., "product-supplier-product"). Assume that multiple paths are sampled from the knowledge graph, such as "electric vehicle battery-battery supplier A-electric vehicle" and "electric vehicle battery-battery supplier B-electric bicycle". These paths are grouped according to the relationship type and the number of hops, for example, all "product-supplier-product" paths with a hop count of 2 are grouped together. The semantic consistency of the nodes in each path group is calculated, such as calculating the average similarity of the word embedding vectors of all nodes in the group to obtain the path group weight. The path groups are ranked according to their importance according to the path group weights, and the top N path groups (e.g., N=3) are selected to construct the core path set of the target product node.

[0137] For each path in the core path set, the conditional probability between adjacent nodes on the path is calculated to obtain the node transfer probability sequence. For example, the transfer probability from "electric vehicle battery" to "battery supplier A" and the transfer probability from "battery supplier A" to "electric vehicle" are calculated. A Markov inference model is constructed based on the node transfer probability sequence. The Monte Carlo method is used to perform random walk sampling on the model to obtain a multi-hop path sampling set. For example, starting from the "electric vehicle battery" node, randomly walk to other nodes according to the transfer probability, and repeat the sampling multiple times to obtain multiple paths. The occurrence frequency of each path in the sampling set is calculated to obtain the multi-hop path probability between the target product node and the supplier node.

[0138] Semantically match the target feature vector with the contextual information of the nodes on each path in the core path set, for example, calculate the similarity between the target feature vector and the word embedding vector of the node on the path to obtain the node importance score. Perform weighted summation on the paths in the core path set based on the node importance score, for example, add up the node importance score of each path to obtain the path importance weight. Perform probability weighted fusion on the path importance weight and the multi-hop path probability, for example, multiply the two to obtain the node association score between the target product node and the supplier node.

[0139] Estimate the distribution of node association scores, such as calculating their mean and standard deviation to obtain the score distribution. Calculate the screening threshold based on the score distribution, such as setting the threshold to the mean plus a standard deviation. The supplier nodes whose node association scores are greater than the screening threshold are used as the preliminary set. Verify the constraints of the supplier nodes in the preliminary set, such as verifying whether they meet the conditions such as "delivery cycle is less than 30 days", "warranty period is not less than 5 years", and "suppliers must pass ISO9001 certification". The verified supplier nodes constitute the candidate supplier node set.

[0140] In this embodiment, the semantic features of the target product are extracted by aligning the distribution of product parameters and constraint parameters, and combined with the graph reasoning mechanism, the core path set with the most business value is screened out based on meta-path sampling and semantic consistency weights. The Markov inference model and Monte Carlo random walk method effectively capture the multi-hop association probability in the supply chain network and fully tap potential supplier resources. The weighted fusion of path importance weights and multi-hop path probabilities not only improves the calculation accuracy of the node association between the target product and the supplier, but also enhances the global perception ability of the model. Through the scoring distribution estimation and dynamic screening threshold, the optimal supplier preliminary set is automatically selected, and combined with the constraint verification mechanism, the business compliance and supply chain stability of the candidate suppliers are ensured, and ultimately efficient, intelligent and reliable supplier recommendations are achieved.

[0141] In an optional implementation, the supplier evaluation data and historical transaction data of the candidate supplier node set are input into a dynamic constraint optimization model, the constraint boundary is updated by online learning, and a multi-objective optimization is performed to obtain a supplier optimization score, and a recommended decision plan including supplier ranking and matching degree is generated based on the supplier optimization score, including:

[0142] Extract supplier evaluation data and historical transaction data of a candidate supplier node set from a dynamic supply chain knowledge graph, construct a temporal cascade attention network to perform multi-scale feature extraction on the quality score, delivery score, and service score in the supplier evaluation data to obtain a comprehensive supplier score, use wavelet transform to perform multi-resolution decomposition on the historical transaction data to obtain trend features and cycle features, construct a transaction fluctuation matrix, a development trend matrix, and a transaction law tensor based on the trend features and cycle features, and obtain a transaction feature vector through tensor decomposition;

[0143] A hierarchical constraint graph structure is constructed according to the transaction feature vector, supply capacity constraint nodes, price range constraint nodes and delivery cycle constraint nodes are established in the hierarchical constraint graph structure, constraint violation data of multiple time windows are collected to construct a time-varying constraint violation graph, structural features of the time-varying constraint violation graph are extracted using a graph convolutional network to obtain a violation degree distribution, and a dynamic constraint boundary update mechanism is constructed based on the violation degree distribution to obtain a constraint condition update sequence;

[0144] Constructing a score maximization objective function based on the supplier comprehensive score, constructing a risk minimization objective function based on the transaction feature vector, and constructing a cost-benefit optimization objective function based on the constraint condition update sequence;

[0145] A multi-objective co-evolutionary optimization framework is constructed, the population is divided into sub-populations for parallel evolution, genes are exchanged between sub-populations through a competitive cooperation mechanism, and adaptive neighborhood search is used to enhance local search capabilities. A dynamic reference vector adjustment strategy is used to maintain population diversity. Multi-objective optimization is performed on the score maximization objective function, risk minimization objective function, and cost-effectiveness optimization objective function. The population is stratified and screened based on dominance level and crowding degree to obtain a multi-objective optimal solution set.

[0146] A multi-head interactive attention network is designed to calculate the dynamic weight of each objective function in the multi-objective optimal solution set, and the dynamic weight and the multi-objective optimal solution set are adaptively fused to obtain the supplier optimization score. The candidate supplier node set is sorted according to the supplier optimization score, and a fuzzy cognitive graph is constructed to calculate the satisfaction degree of each constraint condition and obtain the constraint satisfaction. The supplier optimization score and the constraint satisfaction are input into a deep coupling network to learn the mapping relationship between supplier matching. Based on the sorting results of the supplier optimization score and the supplier matching, a recommended decision plan including supplier sorting and matching is generated.

[0147] For example, first, a set of candidate supplier nodes is extracted from the dynamic supply chain knowledge graph. The knowledge graph stores various information about suppliers, such as the supplier's name, address, contact information, main products, historical transaction data, quality score, delivery score, service score, etc. Suppose we extract five candidate suppliers, named Supplier A, Supplier B, Supplier C, Supplier D, and Supplier E.

[0148] Then, multi-scale feature extraction is performed on the supplier evaluation data. For supplier A, its quality score is 90 points, its delivery score is 85 points, and its service score is 92 points. The temporal cascade attention network is used to extract the features of these three scores at different time scales, such as the score change trend in the short term, the score stability in the long term, etc. These multi-scale features are fused to obtain the comprehensive score of supplier A. The same operation is performed on other suppliers to obtain the comprehensive score of each candidate supplier.

[0149] Next, perform multi-resolution decomposition on the historical transaction data. Assume that the transaction data of supplier A over the past year shows that its transaction amount shows obvious seasonal fluctuations, and there are also some random fluctuations. Wavelet transform is used to decompose the historical transaction data of supplier A into sub-signals of different frequencies, such as low-frequency signals representing long-term trends, medium-frequency signals representing periodic fluctuations, and high-frequency signals representing random fluctuations.

[0150] Based on the decomposed transaction data, the transaction fluctuation matrix, development trend matrix and transaction law tensor are constructed. For example, the transaction fluctuation matrix can reflect the fluctuation of supplier A's transaction volume in different time periods, the development trend matrix can reflect the growth trend of supplier A's transaction volume, and the transaction law tensor can reflect the relationship between supplier A's transaction volume and other factors (such as season, market demand, etc.).

[0151] Through tensor decomposition technology, the transaction pattern tensor is decomposed into multiple transaction feature vectors. These feature vectors represent the core features of the historical transaction data of supplier A. The same operation is performed on other suppliers to obtain the transaction feature vectors of each candidate supplier.

[0152] Construct a hierarchical constraint graph structure based on the transaction feature vector. In the constraint graph, establish supply capacity constraint nodes, price range constraint nodes, and delivery cycle constraint nodes. For example, suppose that supplier A's supply capacity is 1,000 products per month, the price range is 10 to 15 yuan per piece, and the delivery cycle is 7 to 10 days. Collect constraint violation data for multiple time windows, for example, supplier A failed to deliver products on time twice in the past month.

[0153] A time-varying constraint violation graph is constructed, and its structural features are extracted using a graph convolutional network to obtain a violation degree distribution. For example, the violation degree distribution can reflect the violation of supplier A under different constraint conditions. Based on the violation degree distribution, a dynamic constraint boundary update mechanism is constructed to obtain a constraint condition update sequence. For example, if supplier A often fails to deliver products on time, its lead time constraint boundary can be adjusted to a looser range.

[0154] Based on the supplier's comprehensive score, transaction feature vector and constraint update sequence, the score maximization objective function, risk minimization objective function and cost-effectiveness optimization objective function are constructed respectively.

[0155] A multi-objective collaborative evolutionary optimization framework is constructed, the population is divided into sub-populations for parallel evolution, genes are exchanged between sub-populations through a competitive cooperation mechanism, and an adaptive neighborhood search is used to enhance local search capability. A dynamic reference vector adjustment strategy is used to maintain population diversity. Multi-objective optimization is performed on the three objective functions to obtain a multi-objective optimal solution set.

[0156] A multi-head interactive attention network is designed to calculate the dynamic weights of each objective function in the multi-objective optimal solution set, and the dynamic weights and the multi-objective optimal solution set are adaptively fused to obtain the supplier optimization score. Candidate suppliers are ranked according to the supplier optimization score.

[0157] A fuzzy cognitive map is constructed to calculate the satisfaction degree of each constraint condition and obtain the constraint satisfaction degree. The supplier optimization score and constraint satisfaction degree are input into the deep coupling network to learn the mapping relationship of supplier matching degree, and a recommended decision plan including supplier ranking and matching degree is generated based on the ranking results of supplier optimization scores and supplier matching degree.

[0158] In this embodiment, the multi-scale features of supplier evaluation are extracted through a time-series cascade attention network, and the historical transaction data is decomposed in combination with wavelet transform to accurately capture the supplier's transaction trends, cycles and fluctuation characteristics. The hierarchical constraint graph and the time-varying constraint violation graph dynamically update the supply chain constraint boundaries, so that the model can adapt to market changes in real time. The multi-objective co-evolutionary optimization framework fully explores the solution space through population competition and adaptive search mechanisms to obtain the optimal solution set for maximizing supplier scores, minimizing risks and cost-effectiveness. The multi-head interactive attention network dynamically adjusts the weights of each objective to ensure the comprehensiveness and robustness of the recommendation results. The fuzzy cognitive graph and the deep coupling network further learn the mapping relationship of supplier matching, realize the organic integration of supplier ranking and constraint satisfaction, and thus provide a more accurate, reliable and business-constraint-aware supplier recommendation decision solution.

[0159] According to a second aspect of the embodiments of the present invention,

[0160] An electronic device is provided, comprising:

[0161] processor;

[0162] a memory for storing processor-executable instructions;

[0163] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0164] According to a third aspect of the embodiments of the present invention,

[0165] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0166] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A supply chain multidimensional data mining and intelligent recommendation decision-making method based on knowledge graph, characterized in that: include: Supplier evaluation data, product evaluation data and supply and demand transaction data are collected and processed for text segmentation to obtain a word vector sequence, the context association of the word vector sequence is calculated to obtain a semantic feature vector, an initial supply chain knowledge graph is constructed based on the semantic feature vector, the supplier evaluation data, product evaluation data and supply and demand transaction data are used as attribute values ​​of corresponding nodes respectively, the temporal association strength between nodes is calculated to construct a multi-level topological structure, adjacent nodes are iteratively updated based on the multi-level topological structure, and a dynamic supply chain knowledge graph containing the temporal evolution relationship of nodes is generated; Classify the nodes in the dynamic supply chain knowledge graph by type, calculate the information transfer weight between nodes according to the node type, aggregate the node features based on the information transfer weight to obtain a node feature matrix, divide the node feature matrix into multiple time windows according to the time series, calculate the state change of the node in each time window to obtain a time series feature vector, align and fuse the time series feature vector with the node feature matrix to obtain a multidimensional feature representation, calculate the importance weight of each dimensional feature in the multidimensional feature representation, select and reorganize the features according to the importance weight to obtain a fused feature vector, and train the supply chain recommendation model based on the fused feature vector; Extract target product parameters and constraint condition parameters of the target product node from the dynamic supply chain knowledge graph, input them into the supply chain recommendation model, calculate the multi-hop path reachability between the target product node and the supplier node based on the graph reasoning mechanism to obtain the node association score, screen out a set of candidate supplier nodes according to the node association score, input the supplier evaluation data and historical transaction data of the candidate supplier node set into the dynamic constraint optimization model, update the constraint boundary through online learning and perform multi-objective optimization to obtain the supplier optimization score, and generate a recommended decision plan including supplier ranking and matching degree based on the supplier optimization score; Generating a dynamic supply chain knowledge graph containing the temporal evolution relationship of nodes includes: The supplier evaluation data, product evaluation data and supply and demand transaction data are sorted by time series and divided into a forward sequence and a backward sequence, the forward sequence is forward time-series encoded to obtain a forward feature sequence, the backward sequence is reverse time-series encoded to obtain a backward feature sequence, the forward feature sequence and the backward feature sequence are concatenated to obtain supplier node time-series features, product node time-series features and transaction node time-series features, and the supplier node time-series features, product node time-series features and transaction node time-series features are respectively used as node attribute representations of corresponding nodes; Calculating the node attribute variance within the time window based on the node attribute representation, calculating the time sensitivity in combination with the time interval between nodes, and using the time sensitivity to weight the association degree between nodes to obtain the temporal association strength; According to the temporal association strength, the nodes are hierarchically constructed into a multi-level topological structure, the spatial association degree between the nodes in the same layer in the multi-level topological structure is calculated to obtain a local association score, the temporal association degree between the nodes across layers is calculated to obtain a long-range association score, and the local association score and the long-range association score are integrated to obtain a comprehensive association strength between the nodes; The historical state information of the node is stored, and the relevant historical state information is retrieved based on the current state of the node to obtain the node memory feature, and the states of the adjacent nodes are weighted and aggregated according to the comprehensive correlation strength between the nodes to obtain the node neighbor feature, and the node memory feature and the node neighbor feature are iteratively updated to obtain a new node state, and the stored historical state information is updated; The updated node status is used as the dynamic feature representation of the node, and the node connection relationship is constructed according to the comprehensive correlation strength between the nodes to generate a dynamic supply chain knowledge graph containing the time-series evolution relationship of the nodes.

2. The method according to claim 1, characterized in that Collecting supplier evaluation data, product evaluation data and supply and demand transaction data and performing text segmentation processing to obtain a word vector sequence, calculating the context association of the word vector sequence to obtain a semantic feature vector, and constructing an initial supply chain knowledge graph based on the semantic feature vector includes: Perform word segmentation preprocessing on supplier evaluation data, product evaluation data and supply and demand transaction data to obtain a word vector sequence, calculate the frequency of occurrence of words in the word vector sequence in a single document and the distribution frequency in all document sets, construct a word weight calculation model to obtain a word weight matrix, extract a feature word sequence based on the word weight matrix, use a conditional random field model to annotate the feature word sequence to obtain an annotated word sequence, construct the annotated word sequence as an initial word library, construct a word dependency tree based on the initial word library to extract a syntactic dependency matrix between words, and perform hierarchical clustering on the initial word library according to the syntactic dependency matrix to obtain a dynamic word library; Constructing a word co-occurrence matrix for the words in the dynamic vocabulary, calculating a word semantic similarity matrix based on the word co-occurrence matrix, using a bidirectional long short-term memory network to process the word semantic similarity matrix to extract context features to obtain a semantic representation vector group, constructing a time-aware mechanism to calculate a temporal weight coefficient, weighting the vectors in the semantic representation vector group according to the temporal weight coefficient to obtain a weighted vector group, and performing temporal fusion on the weighted vector group to obtain a semantic feature vector; The supplier node, product node and transaction node are used as graph network vertices, the semantic feature vector is mapped into a node feature matrix, the node feature matrix is ​​processed and calculated by a graph attention network to obtain a multi-dimensional association probability matrix between nodes, a probability transfer matrix is ​​generated based on the multi-dimensional association probability matrix, and the node features are updated under the guidance of the probability transfer matrix through a message passing mechanism to obtain a node hidden layer feature matrix; The matching coefficient matrix between nodes is calculated based on the node hidden layer feature matrix, wherein the matching coefficient matrix includes the supply and demand matching coefficient between the supplier node and the product node, the similarity coefficient between the product nodes, and the correlation coefficient between the supplier node and the transaction node. The matching coefficient matrix is ​​used as the initial weight matrix of the graph edge, and the random walk method is used to perform multiple rounds of iterative optimization on the initial weight matrix to obtain the final edge weight matrix. The initial supply chain knowledge graph is constructed based on the final edge weight matrix.

3. The method according to claim 1, characterized in that Classifying the nodes in the dynamic supply chain knowledge graph, calculating the information transfer weight between nodes according to the node type, aggregating the node features based on the information transfer weight to obtain a node feature matrix, dividing the node feature matrix into multiple time windows according to the time series, and calculating the state change of the node in each time window to obtain the time series feature vector includes: Obtain the global statistical features and local structural features of the nodes in the dynamic supply chain knowledge graph, construct a multidimensional semantic space including business function dimensions, resource occupation dimensions and timing characteristics dimensions based on the global statistical features and local structural features, map the nodes to the multidimensional semantic space to obtain an initial semantic vector, use a multilayer perceptron to fuse the global statistical features and local structural features to obtain a dimensional weight coefficient, multiply the dimensional weight coefficient by the corresponding dimensional feature of the initial semantic vector to obtain a weighted semantic vector, and normalize the weighted semantic vector to obtain a node type representation; Constructing a query vector and a key vector based on the node type representation, calculating the forward similarity between the query vector and the key vector between the node pair to obtain a forward attention score, calculating the reverse similarity between the key vector and the query vector between the node pair to obtain a reverse attention score, concatenating the forward attention score and the reverse attention score and inputting them into a gating unit to obtain a gating coefficient, and using the gating coefficient to perform a weighted combination of the forward attention score and the reverse attention score to obtain an information transfer weight between nodes; Construct a multi-layer feature aggregation network, perform weighted aggregation of the features of adjacent nodes in each layer of the network according to the information transmission weights between the nodes, and map the aggregated features to a new feature space through an inter-layer conversion matrix to obtain the node representation of the current layer, superimpose the node representations of each layer to obtain a node feature matrix, and perform sequence division of the node feature matrix according to a preset time window size to obtain a time series feature sequence; For each time window in the temporal feature sequence, a short-term convolution kernel, a medium-term convolution kernel and a long-term convolution kernel are used to perform convolution operations on the node feature matrix in the window to obtain short-term feature sequences, medium-term feature sequences and long-term feature sequences representing patterns of different time scales; an attention query matrix is ​​constructed based on the short-term feature sequence, medium-term feature sequence and long-term feature sequence, the similarity between the feature sequence and the attention query matrix is ​​calculated to obtain feature weights, and the feature weights are used to perform weighted combination of feature sequences of different scales to obtain a temporal feature vector.

4. The method according to claim 1, characterized in that: Aligning and fusing the time series feature vector with the node feature matrix to obtain a multidimensional feature representation, calculating the importance weight of each dimensional feature in the multidimensional feature representation, selecting and reorganizing the features according to the importance weight to obtain a fused feature vector, and training a supply chain recommendation model based on the fused feature vector includes: A bidirectional feature mapping network is constructed, and the input time series feature vector and node feature matrix are respectively nonlinearly transformed by the bidirectional feature mapping network to obtain projection features, the projection features with the same time series position are constructed as positive sample pairs, and the projection features and randomly selected projection features with different time series positions are constructed as negative sample pairs, and the feature similarity of the positive sample pair is maximized and the feature similarity of the negative sample pair is minimized by optimizing the contrast learning objective function to obtain alignment features; Performing a linear transformation on the multidimensional feature representation to obtain a query matrix, a key matrix, and a value matrix, calculating the dot product of the query matrix and the key matrix to obtain an attention score, normalizing the attention score and multiplying it with the value matrix to obtain a feature importance weight matrix, applying sparse constraints and orthogonal constraints to the feature importance weight matrix to obtain feature weights; According to the feature weights, each dimension of the multidimensional feature representation is sorted and screened to obtain important features, and the important features and the multidimensional feature representation are reconstructed by residual connection to obtain a fused feature vector; Constraints are extracted from the supply chain system and converted into constraint vectors through a constraint encoding network. The fused feature vector and the constraint vector are inner-producted to obtain a recommendation score. A recommendation loss term is constructed based on the recommendation score. A constraint violation penalty term is constructed based on the degree of constraint satisfaction. A smoothing regularization term is added to construct a joint loss function. The joint loss function is optimized and trained using the gradient descent method to obtain a supply chain recommendation model.

5. The method according to claim 1, characterized in that Target product parameters and constraint condition parameters of the target product node are extracted from the dynamic supply chain knowledge graph and input into the supply chain recommendation model. The multi-hop path reachability between the target product node and the supplier node is calculated based on the graph reasoning mechanism to obtain the node association score. The candidate supplier node set is screened out according to the node association score, including: Extract target product parameters and constraint condition parameters of the target product node from the dynamic supply chain knowledge graph, map the target product parameters to distribution vectors in the product semantic feature space, sample and generate target product features from the distribution vectors, decompose the constraint condition parameters to obtain a constraint condition feature sequence, construct a constraint condition vector according to the constraint condition feature sequence, and align the target product features with the constraint condition vector for probability distribution to obtain a target feature vector; Construct a graph reasoning mechanism, take the target product node in the dynamic supply chain knowledge graph as the starting point, perform group path sampling based on the predefined meta-path pattern, group the sampled paths according to the relationship type and the number of hops to obtain a path group set, calculate the semantic consistency of the nodes in each path group to obtain the path group weight, sort the path group set by importance according to the path group weight, and select the top N path groups to construct the core path set of the target product node; For each path in the core path set, the conditional probability between adjacent nodes on the path is calculated to obtain a node transfer probability sequence, a Markov inference model is constructed based on the node transfer probability sequence, a Monte Carlo method is used to perform random walk sampling on the Markov inference model to obtain a multi-hop path sampling set, and the occurrence frequency of each path in the multi-hop path sampling set is calculated to obtain a multi-hop path probability between a target product node and a supplier node; Perform semantic matching on the target feature vector and the context information of the nodes on each path in the core path set to obtain a node importance score, perform weighted summation on the paths in the core path set based on the node importance score to obtain a path importance weight, and perform probability weighted fusion on the path importance weight and the multi-hop path probability to obtain a node association score between the target product node and the supplier node; The node association scores are distributed and estimated to obtain a score distribution. A screening threshold is calculated based on the score distribution. The supplier nodes whose node association scores are greater than the screening threshold are taken as a preliminary set. Constraints are verified for the supplier nodes in the preliminary set, and the supplier nodes that pass the verification constitute a candidate supplier node set.

6. The method according to claim 1, characterized in that The supplier evaluation data and historical transaction data of the candidate supplier node set are input into the dynamic constraint optimization model, the constraint boundary is updated by online learning, and a multi-objective optimization is performed to obtain the supplier optimization score. Based on the supplier optimization score, a recommended decision plan including supplier ranking and matching degree is generated, including: Extract supplier evaluation data and historical transaction data of a candidate supplier node set from a dynamic supply chain knowledge graph, construct a temporal cascade attention network to perform multi-scale feature extraction on the quality score, delivery score, and service score in the supplier evaluation data to obtain a comprehensive supplier score, use wavelet transform to perform multi-resolution decomposition on the historical transaction data to obtain trend features and cycle features, construct a transaction fluctuation matrix, a development trend matrix, and a transaction law tensor based on the trend features and cycle features, and obtain a transaction feature vector through tensor decomposition; A hierarchical constraint graph structure is constructed according to the transaction feature vector, supply capacity constraint nodes, price range constraint nodes and delivery cycle constraint nodes are established in the hierarchical constraint graph structure, constraint violation data of multiple time windows are collected to construct a time-varying constraint violation graph, structural features of the time-varying constraint violation graph are extracted using a graph convolutional network to obtain a violation degree distribution, and a dynamic constraint boundary update mechanism is constructed based on the violation degree distribution to obtain a constraint condition update sequence; Constructing a score maximization objective function based on the supplier comprehensive score, constructing a risk minimization objective function based on the transaction feature vector, and constructing a cost-benefit optimization objective function based on the constraint condition update sequence; A multi-objective co-evolutionary optimization framework is constructed, the population is divided into sub-populations for parallel evolution, genes are exchanged between sub-populations through a competitive cooperation mechanism, and adaptive neighborhood search is used to enhance local search capabilities. A dynamic reference vector adjustment strategy is used to maintain population diversity. Multi-objective optimization is performed on the score maximization objective function, risk minimization objective function, and cost-effectiveness optimization objective function. The population is stratified and screened based on dominance level and crowding degree to obtain a multi-objective optimal solution set. A multi-head interactive attention network is designed to calculate the dynamic weight of each objective function in the multi-objective optimal solution set, and the dynamic weight and the multi-objective optimal solution set are adaptively fused to obtain the supplier optimization score. The candidate supplier node set is sorted according to the supplier optimization score, and a fuzzy cognitive graph is constructed to calculate the satisfaction degree of each constraint condition and obtain the constraint satisfaction. The supplier optimization score and the constraint satisfaction are input into a deep coupling network to learn the mapping relationship between supplier matching. Based on the sorting results of the supplier optimization score and the supplier matching, a recommended decision plan including supplier sorting and matching is generated.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images