Potential supplier recommendation method based on knowledge graph enhanced relational graph convolutional network

Abstract

The invention discloses a potential supplier recommendation method based on a knowledge graph enhanced relational graph convolutional network, and relates to the technical field of convolutional networks, and the method comprises the steps: firstly collecting supply chain data, carrying out the preprocessing and rule matching, screening entities and relationships, carrying out the standardization, and constructing a supply chain knowledge graph; processing the knowledge graph by using a relational graph convolutional network to obtain an enhanced model; training the model and predicting the cooperation possibility of the supplier and the customer; and finally, if the evaluation result does not reach the standard, training is repeated until the expectation is met. According to the method, a supply chain knowledge graph is constructed to capture a multivariate relation, a relation graph convolutional network is used for modeling a multi-relation structure, entity embedding containing structures and semantics is generated, the problem that a traditional system processes a complex relation is solved, the effect is better than that of models such as the graph convolutional network, over-fitting is avoided by combining a loss function and regularization, entity representation is enriched by the knowledge graph, and the method has a good application prospect. And multi-relation modeling improves interpretability, provides insight for decision making, and assists in improving the toughness of the supply chain.

Description

Technical Field

[0001] The present invention relates to the field of convolutional network technology, and in particular to a potential supplier recommendation method based on a knowledge graph enhanced relational graph convolutional network. Background Art

[0002] A deeply bound industrial collaboration relationship has been formed between suppliers and customers, which combines the rigor of a closed supply and demand loop with the flexibility of dynamic collaboration. As core participants in the upstream of the industrial chain, suppliers provide customers with full-chain support from key components to raw materials and technical solutions. The stability of their product quality, the accuracy of their delivery cycles and their cost control capabilities directly determine the production rhythm, product performance and even market competitiveness of downstream customers. Customers lead the direction of cooperation through order planning, technical standard output, demand feedback, etc., deepen collaboration through long-term agreements and joint R&D mechanisms, and discuss alternative solutions with suppliers in the face of supply chain risks. This two-way interaction not only lays a solid foundation for the efficient operation of the industrial chain, but also promotes the collaborative evolution of both supply and demand sides in technological iteration and market competition, and jointly maintains the dynamic balance of the industrial ecology.

[0003] For example, Chinese patent publication number: CN114817712B provides a project recommendation method based on multi-task learning and knowledge graph enhancement, including books, movies, music, and commodities. Compared with traditional project recommendation methods, the present invention introduces knowledge graphs to alleviate problems such as data sparsity and low precision. The present invention designs a recommendation model EMKR, which uses a multi-task learning shared unit association recommendation module and a knowledge graph embedding module. In the recommendation module, a two-part modeling strategy is used, the attention mechanism is used to capture user behavior patterns, and auxiliary vectors are introduced to expand the potential expression ability of users and projects; in the knowledge graph embedding module, the importance of different relationships is distinguished, and the graph convolutional network is used to mine the rich multi-relationship semantics of the knowledge graph. Finally, EMKR is trained alternately, and the trained model is used for recommendation. The project recommendation method disclosed by the present invention has good performance, especially in sparse recommendation scenarios, the effect is significantly improved.

[0004] However, the above solution does not take into account the high complexity and diversity of entity relationships in modern supply chain networks. Traditional supplier recommendation systems are difficult to adapt to this complex pattern. They can neither effectively handle the intricate interactive relationships and multiple entity associations in the network, nor accurately capture the deep connections between suppliers, customers, products and other entities. This directly leads to poor performance in scenarios where timely and accurate recommendations of alternative suppliers are needed to ensure stable operations, and it is difficult to meet the actual needs of enterprises to improve supply chain resilience and reduce interruption risks. Summary of the Invention

[0005] Technical problems solved In response to the shortcomings of the existing technology, the present invention provides a potential supplier recommendation method based on a knowledge graph enhanced relational graph convolutional network, which solves the problem that traditional supplier recommendation systems are difficult to adapt to the high complexity and diversity of entity relationships in modern supply chains, cannot accurately capture the deep connections between entities, and are weak in timely recommending alternative suppliers to ensure operational stability, improve supply chain resilience and reduce interruption risks.

[0006] Technical Solution To achieve the above objectives, the present invention is implemented through the following technical solutions: a potential supplier recommendation method based on a knowledge graph-enhanced relational graph convolutional network, comprising the following specific steps: Step 1: Collect supply chain data and pre-process it, filter out entity data and entity relationship data through rule matching, and construct a supply chain knowledge graph after standardization; Step 2: Perform graph convolution operation on the supply chain knowledge graph through the relational graph convolutional network, first set the basic elements, the basic elements include nodes In the Feature embedding representation of the layer, neighbor nodes In the The feature embedding representation of the layer, the entity relationship type set, the node relationship, the relationship-specific matrix, the self-loop conversion matrix and the normalization factor are then processed, and the neighbor information is processed according to the relationship type, and the neighbor contribution under a specific relationship in the entity relationship type set is extracted. Then, the neighbor information of all relationships is aggregated, and the node information itself is continued to be processed. Finally, the information is fused and activated to obtain the knowledge graph enhanced relationship graph convolutional network model; Step three: Optimize and train the knowledge graph enhanced relationship graph convolutional network model, and predict the possibility of cooperation between suppliers and customers based on the knowledge graph enhanced relationship graph convolutional network model to obtain the prediction result; Step four: Evaluate the prediction result. If the prediction result meets the expectation, the supplier is recommended and the process ends. Otherwise, return to step three until the prediction result meets the expectation.

[0007] Furthermore, the specific construction method of the supply chain knowledge graph is as follows: statistics are performed on entity data and entity relationship data to obtain all entity sets and all entity relationship sets, statistics are performed on the types of entities in all entity sets to obtain entity type sets, statistics are performed on the types of entity relationships in all entity relationship sets to obtain entity relationship type sets, and a certain entity in all entity sets, a certain entity relationship in all entity relationship sets, the entity type of a certain entity in all entity sets, and the entity relationship type of a certain entity relationship in all entity relationship sets are defined; based on all entity sets, all entity relationship sets, entity type sets, entity relationship type sets, a certain entity in all entity sets, a certain entity relationship in all entity relationship sets, the entity type of a certain entity in all entity sets, and the entity relationship type of a certain entity relationship in all entity relationship sets, construct an ontology structure of entities and relationship types, which is recorded as a supply chain knowledge graph.

[0008] Furthermore, the specific method of obtaining the knowledge graph enhanced relational graph convolutional network model is as follows: ;in, Representation node In the The feature embedding representation of the layer, Through node relationships and The set of connected adjacent nodes, Representation node In the The feature embedding representation of the layer, Representation node In the The feature embedding representation of the layer, Represents a set of entity relationship types, each node relationship All with unique relationship exclusive matrix Related, is the normalization function, is the self-loop weight matrix, function Represents a nonlinear activation function, usually the rectified linear unit ReLU.

[0009] Furthermore, the specific steps of processing neighbor information by relationship type are: for each neighbor node under the node relationship, use the relationship-specific matrix corresponding to a specific relationship in the entity relationship type set to process the node In the The feature embedding of the layer is linearly transformed to obtain ,Will The product calculation is performed with the normalization function, and the normalized information of all neighbor nodes under the node relationship is summed up to obtain the total contribution of the neighbor nodes to the node under a specific relationship in the entity relationship type set.

[0010] Furthermore, the specific step of aggregating neighbor information of all relationships is: summing up the total contribution of neighbor nodes to the node under a specific relationship in the entity relationship type set to obtain the total information transmitted to the node by all neighbor nodes through different relationships.

[0011] Furthermore, the specific steps of processing the node's own information are: Layer feature nodes In the The feature embedding representation of the layer is linearly transformed using the self-loop transformation matrix to obtain .

[0012] Furthermore, the specific steps of fusing information and activating are: combining the total information transmitted to the node by all neighbor nodes through different relationships with Perform addition calculation to obtain the total information of inactivation, apply activation function to the total information of inactivation, introduce nonlinear transformation, and obtain node In the The feature embedding representation of the layer.

[0013] Furthermore, the specific steps for optimizing and training the knowledge graph enhanced relational graph convolutional network model are as follows: using a set of labeled supplier-customer interaction data, including positive samples and negative samples, i.e., samples of real cooperation and samples of non-real cooperation. The learning goal is to minimize the difference between the predicted score and the true label through the binary cross entropy loss function. After the training is completed, the model outputs a probability score to quantify the possibility of cooperation between each pair of supplier-customers. By ranking these scores, the system generates a personalized supplier recommendation list, in which suppliers with higher rankings have a greater chance of becoming suitable partners.

[0014] Furthermore, the specific method of the prediction result is as follows: ;in, Represents the prediction results, represents the activation function, Represents the customer entity The relational graph convolutional network model enhanced by knowledge graph The final feature embedding obtained after layer processing, Represents a supplier entity The relational graph convolutional network model enhanced by knowledge graph The final feature embedding obtained after layer processing, Represents a learnable weight matrix for subsequent model training.

[0015] Furthermore, the specific method of evaluating the prediction results is: using four indicators to evaluate the model, namely the area under the receiver operating characteristic curve, precision rate, overall accuracy rate and recall rate, setting the sample k and sample set for evaluation, sample k belongs to the sample set, and the sample set includes 20, 50 and 100.

[0016] Beneficial effects Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: 1. By constructing a supply chain knowledge graph to capture the multi-dimensional relationships between entities such as suppliers, customers, and products, and using a relational graph convolutional network to model the multi-relationship graph structure, the relational graph convolutional network distinguishes different types of entity relationships through a relationship-specific matrix. Combined with normalized aggregation and self-loop feature retention, it generates entity embeddings rich in structural and semantic information, solving the problem that traditional recommendation systems have difficulty handling complex interactive relationships.

[0017] 2. By combining the binary cross-entropy loss function with L2 regularization constraint weights, overfitting is effectively avoided and the model's predictive stability for unseen supplier-customer pairs is improved. At the same time, the structured semantic information provided by the knowledge graph enhances the richness of entity representation, and the multi-relationship modeling mechanism enables the model to explore potential associations from multiple dimensions, with strong interpretability. Case studies show that the model can reveal the key influencing factors of recommendation results, provide decision makers with actionable insights, and help improve supply chain resilience and risk response capabilities.

[0018] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic diagram of the ontology structure of entities and relationship types of the present invention.

[0020] Figure 2 The present invention is a graph showing the increasing trend of the area under the receiver operating characteristic curve, the precision rate, and the overall accuracy rate with the number of training iterations.

[0021] Figure 3 The present invention shows a steadily increasing trend graph of the recall rate corresponding to different sample k values ​​as the number of training iterations increases. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0023] It should be noted that, in this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include," "comprise," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0024] Example 1: like Figure 1 As shown, the embodiment of the present invention provides a potential supplier recommendation method based on a knowledge graph enhanced relational graph convolutional network, which includes the following specific steps: Step 1: Collect supply chain data and perform data cleansing. Supply chain data is crawled from the supply chain database. Entity data and entity relationship data are filtered out through rule matching, that is, entity data and entity relationship data are filtered out through pre-set data fixed formats and data associations. Entity data includes suppliers, customers, etc. Entity relationship data includes the relationship between suppliers and customers, the relationship between suppliers and product categories, etc. Supply chain data is cleaned to avoid data redundancy and improve the quality of supply chain data. After standardization, the dimensional differences of the supply chain data are eliminated. A supply chain knowledge graph is constructed based on the entity data and entity relationship data, that is, a heterogeneous, directed, and static network. Step 2: Perform graph convolution on the supply chain knowledge graph through the relational graph convolution network, which can effectively capture the heterogeneous relationships between entities in the supply chain knowledge graph, enhance entity association modeling, and obtain a knowledge graph enhanced relational graph convolution network model; Step 3: Obtain the embedded representations of all entities in the supply chain knowledge graph based on the knowledge graph-enhanced relational graph convolutional network model, predict the possibility of cooperation between suppliers and customers, obtain the prediction results, and optimize the knowledge graph-enhanced relational graph convolutional network model for training; Step 4: Evaluate the prediction results. If the prediction results meet expectations, the process ends. Otherwise, return to step 3 until the prediction results meet expectations.

[0025] Example 2 differs from Example 1 in that: The specific construction method of the supply chain knowledge graph is as follows: Entity data and entity relationship data are statistically analyzed to obtain all entity sets and all entity relationship sets. All entity sets include entities such as suppliers, customers, countries, and product capabilities. All entity relationship sets include the relationships between entities. The types of entities in all entity sets are statistically analyzed to obtain an entity type set. The types of entity relationships in all entity relationship sets are statistically analyzed to obtain an entity relationship type set. Define an entity in all entity sets, an entity relationship in all entity relationship sets, an entity type of an entity in all entity sets, and an entity relationship type of an entity relationship in all entity relationship sets. Construct an ontology structure of entities and relationship types based on all entity sets, all entity relationship sets, entity type sets, entity relationship type sets, an entity in all entity sets, an entity relationship in all entity relationship sets, an entity type of an entity in all entity sets, and an entity relationship type of an entity relationship in all entity relationship sets, which is recorded as a supply chain knowledge graph. ; in, Represents the ontology structure of entity and relationship types, that is, to establish connections between entities of all entity types and entity relationships of corresponding entity relationship types. Represents an entity in the set of all entities, Represents an entity relationship in the set of all entity relationships, Represents the entity type of an entity in the collection of all entities, The entity relationship type that represents an entity relationship in the set of all entity relationships. Represents the set of all entities, Represents the set of all entity relationships, Represents a collection of entity types, Represents a set of entity relationship types, Indicates that an entity in the set of all entities is classified as the entity type of an entity in the set of all entities, that is, this entity is an object of a certain entity type. Indicates that an entity relationship in the set of all entity relationships is classified as an entity relationship type of an entity relationship in the set of all entity relationships, that is, this entity relationship is an object of a certain entity relationship type.

[0026] The specific method of obtaining the knowledge graph enhanced relational graph convolutional network model is as follows: First, set up the basic elements to provide the necessary basic information and clear processing scope for subsequent calculations. The basic elements include nodes In the Feature embedding representation of the layer, neighbor nodes In the The feature embedding representation of the layer, the entity relationship type set, the node relationship, the relationship-specific matrix, the self-loop conversion matrix and the normalization factor are then processed. The neighbor information is then processed according to the relationship type. The neighbor information under each relationship type is transformed and normalized in a targeted manner. The neighbor contribution under a specific relationship in the entity relationship type set is extracted. The neighbor information of all relationships is aggregated, and the neighbor contribution under all relationship types is accumulated to obtain the overall neighbor information sum. The node's own information is then processed and its own features are transformed to ensure that its core information is retained during the update. Finally, the information is fused and activated. The total neighbor information and the node's own information are fused and processed by the activation function to obtain the feature representation of the next layer of the node. ; in, Representation node In the The feature embedding representation of the layer is represented by the neighbor nodes The recursive aggregation of node relationships accumulates structural and semantic information from the supply chain knowledge graph. Through node relationships and The set of connected adjacent nodes, Representation node In the The feature embedding representation of the layer, Representation node In the The feature embedding representation of the layer, Represents a set of entity relationship types, each node relationship All with unique relationship exclusive matrix Correlation,regulates the information flow based on relational semantics, ensuring that different types of node relations contribute differently to node updates, thus enabling the model to distinguish them. is a normalization function that considers the node relationship midpoint With neighboring nodes The degree of the nodes is used to normalize the aggregation, thus ensuring balanced aggregation regardless of the size of the neighbor nodes. Is the self-loop weight matrix, which allows the node to retain its own characteristics during the update process. Function Represents a nonlinear activation function, usually a rectified linear unit (ReLU), which enables the model to capture complex nonlinear patterns in the supply chain knowledge graph.

[0027] The specific steps for processing neighbor information by relationship type are: For each neighbor node under the node relationship, use the relationship-specific matrix corresponding to a specific relationship in the entity relationship type set to calculate the node In the The feature embedding of the layer is linearly transformed to obtain , to ensure that the information of different relationships can be effectively distinguished by the model, The product calculation is performed with the normalization function to avoid the information of highly connected nodes being over-amplified or the information of low-connected nodes being ignored. The normalized information of all neighbor nodes under the node relationship is summed up to obtain the total contribution of neighbor nodes to the node under a specific relationship in the entity relationship type set.

[0028] The specific steps to aggregate neighbor information of all relationships are: The total contribution of neighbor nodes to the node under a specific relationship in the entity relationship type set is summed up to obtain the total information transmitted to the node by all neighbor nodes through different relationships, that is, the neighbor node information scattered in different relationship types is integrated into a unified global neighbor node influence.

[0029] The specific steps for processing node information are: For its Layer feature nodes In the The feature embedding representation of the layer is linearly transformed using the self-loop transformation matrix to obtain , to avoid its own information being overwhelmed by neighbor information.

[0030] The specific steps of information fusion and activation are: The total information transmitted to the node by all neighbor nodes through different relationships is Perform sum calculation to obtain the total information of inactivation, apply activation function to the total information of inactivation, introduce nonlinear transformation, and obtain node In the The feature embedding representation of the layer.

[0031] The specific method of predicting the results is as follows: ; in, Represents the prediction results, represents the activation function, Represents the customer entity The relational graph convolutional network model enhanced by knowledge graph The final feature embedding obtained after layer processing, Represents a supplier entity The relational graph convolutional network model enhanced by knowledge graph The final feature embedding obtained after layer processing, Represents a learnable weight matrix for subsequent model training.

[0032] The specific steps for optimizing and training the knowledge graph enhanced relational graph convolutional network model are as follows: To train the model, the present invention uses a set of labeled supplier-customer interaction data, including positive samples and negative samples, that is, samples of real cooperation and samples of non-real cooperation. The learning goal is to minimize the difference between the predicted score and the true label through the binary cross-entropy loss function. After the training is completed, the model outputs a probability score to quantify the possibility of cooperation between each pair of supplier-customer. By ranking these scores, the system generates a personalized supplier recommendation list, where suppliers with higher rankings have a greater chance of becoming suitable partners.

[0033] The specific method for optimizing and training the knowledge graph enhanced relational graph convolutional network model is as follows: ; in, represents the loss function, represents the total number of training samples, It is between 0 and 1, indicating whether the supplier and the customer are truly cooperating. Represents the prediction results, represents the regularization coefficient, Represents the sum of squares of all elements in the relationship-specific matrix to prevent the model from over-relying on a certain weight.

[0034] The specific method for evaluating the prediction results is as follows: Four metrics are used to evaluate the model, namely the area under the receiver operating characteristic curve, precision, overall accuracy, and recall. Based on previous literature, the value of sample k is set to 20, 50, and 100 for evaluation. The following is a detailed explanation of each metric. The area under the receiver operating characteristic curve is a score between 0 and 1, which measures the performance of the model relative to a random classifier. A score of 1 indicates that the model performs very well and has strong discrimination ability. The precision rate is also a score between 0 and 1, which reflects the proportion of correctly predicted positive examples among all samples predicted as positive examples. A score of 1 means that the model does not produce false positives, that is, all predicted positive examples are correct. The overall accuracy rate ranges from 0 to 1, which measures the proportion of correct predictions among all predictions. A higher overall accuracy rate indicates that the model performs well on both positive and negative samples. The recall rate is an indicator that measures the ability of the model to identify positive samples. It is calculated as the ratio of the number of true positive samples to the total number of all relevant samples in the test set.

[0035] Application Simulation: like Figure 2-Figure 3As shown in the figure, the relational graph convolutional network model enhanced by knowledge graph is implemented by PyTorch Geometric library. All experiments are conducted on a desktop computer equipped with Intel Core i7-10750H 2.6GHz processor (16 GB memory) and NVIDIA GeForce RTX 4060 graphics card (8 GB video memory). The dataset contains 869 customers, 11333 suppliers and 190052 relationships. The training / test set is divided into 3:1, and the positive and negative samples are balanced with a negative sample sampling rate of 1. In order to verify the adaptation advantage of the relational graph convolutional network to multi-relational graphs, the relational graph convolutional network model enhanced by knowledge graph is compared with the knowledge graph + graph convolutional network / graph attention network model. At the same time, the relational graph convolutional network enhanced by knowledge graph and graph neural network baseline are included to highlight the gain of knowledge graph and multi-relational modeling. Experimental control: the model is trained for 1000 rounds and repeated 10 times with random initialization to ensure stability; after hyperparameter tuning, the relational graph convolutional network dropout rate is 0.2, the learning rate , adapted to 6 types of relationships, L2 regularization coefficient , gradient clipping threshold 1.0, node input dimension 10; The knowledge graph-enhanced relational graph convolutional network model performed well on the training set, with an area under the receiver operating characteristic curve of nearly 0.9, a precision rate and an overall accuracy rate exceeding 0.8, and a recall rate of approximately 0.83. All indicators were significantly improved. The comparison of the test set showed that the model with the introduction of the knowledge graph was better than the corresponding model without the introduction of the knowledge graph, because the knowledge graph provided structured semantic information to enrich the node representation. The knowledge graph-enhanced relational graph convolutional network model performed best in comparison with the graph convolutional network and graph attention network with the introduction of the knowledge graph, with an area under the receiver operating characteristic curve of 0.8941, an overall accuracy of 0.8193, and an excellent recall rate indicator. This is due to its ability to accurately model multiple relations, distinguish edge types and utilize contextual information, which is better than the graph convolutional network and graph attention network that have limitations in processing multiple relations.

[0036] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A potential supplier recommendation method based on a knowledge graph-enhanced relational graph convolutional network, characterized by: The specific steps include: Step 1: Collect and pre-process supply chain data, filter out entity data and entity relationship data through rule matching, standardize them, and build a supply chain knowledge graph; Step 2: Perform graph convolution on the supply chain knowledge graph through the relational graph convolution network. First, set the basic elements, which include nodes In the Feature embedding representation of the layer, neighbor nodes In the The feature embedding representation of the layer, the entity relationship type set, the node relationship, the relationship-specific matrix, the self-loop conversion matrix and the normalization factor are then processed. The neighbor information is then processed according to the relationship type, and the neighbor contribution of a specific relationship in the entity relationship type set is extracted. The neighbor information of all relationships is then aggregated, and the node information itself is further processed. Finally, the information is fused and activated to obtain a knowledge graph enhanced relational graph convolutional network model. Step 3: Optimize and train the knowledge graph-enhanced relational graph convolutional network model, and predict the possibility of cooperation between suppliers and customers based on the knowledge graph-enhanced relational graph convolutional network model to obtain prediction results; Step 4: Evaluate the forecast results. If the forecast results meet expectations, recommend suppliers and end the process. Otherwise, return to step 3 until the forecast results meet expectations.

2. The method for recommending potential suppliers based on a knowledge graph-enhanced relational graph convolutional network according to claim 1, characterized in that: The specific construction method of the supply chain knowledge graph is as follows: Count the entity data and entity relationship data to obtain all entity sets and all entity relationship sets, count the types of entities in all entity sets to obtain an entity type set, count the types of entity relationships in all entity relationship sets to obtain an entity relationship type set, define an entity in all entity sets, an entity relationship in all entity relationship sets, the entity type of an entity in all entity sets, and the entity relationship type of an entity relationship in all entity relationship sets; An ontology structure of entities and relationship types is constructed based on all entity sets, all entity relationship sets, entity type sets, entity relationship type sets, an entity in all entity sets, an entity relationship in all entity relationship sets, the entity type of an entity in all entity sets, and the entity relationship type of an entity relationship in all entity relationship sets, which is recorded as a supply chain knowledge graph.

3. The method for recommending potential suppliers based on a knowledge graph-enhanced relational graph convolutional network according to claim 1, characterized in that: The specific method of obtaining the knowledge graph enhanced relational graph convolutional network model is as follows: ； in, Representation node In the The feature embedding representation of the layer, Through node relationships and The set of connected adjacent nodes, Representation node In the The feature embedding representation of the layer, Representation node In the The feature embedding representation of the layer, Represents a set of entity relationship types, each node relationship All have unique relationship with the exclusive matrix Related, is the normalization function, is the self-loop weight matrix, function Represents a nonlinear activation function, usually the rectified linear unit ReLU.

4. The method for recommending potential suppliers based on a knowledge graph-enhanced relational graph convolutional network according to claim 3 is characterized by: The specific steps of processing neighbor information according to relationship type are: For each neighbor node under the node relationship, use the relationship-specific matrix corresponding to a specific relationship in the entity relationship type set to calculate the node In the The feature embedding of the layer is linearly transformed to obtain ,Will The product calculation is performed with the normalization function, and the normalized information of all neighbor nodes under the node relationship is summed up to obtain the total contribution of the neighbor nodes to the node under a specific relationship in the entity relationship type set.

5. The method for recommending potential suppliers based on a knowledge graph-enhanced relational graph convolutional network according to claim 3 is characterized by: The specific steps of aggregating neighbor information of all relationships are: The total contribution of neighbor nodes to the node under a specific relationship in the entity relationship type set is summed up to obtain the total information transmitted to the node by all neighbor nodes through different relationships.

6. The method for recommending potential suppliers based on a knowledge graph-enhanced relational graph convolutional network according to claim 3, characterized in that: The specific steps of processing the node's own information are: For its Layer feature nodes In the The feature embedding representation of the layer is linearly transformed using the self-loop transformation matrix to obtain .

7. The method for recommending potential suppliers based on a knowledge graph-enhanced relational graph convolutional network according to claim 3, characterized in that: The specific steps of fusing information and activating are: The total information transmitted to the node by all neighbor nodes through different relationships is Perform addition calculation to obtain the total information of inactivation, apply activation function to the total information of inactivation, introduce nonlinear transformation, and obtain node In the The feature embedding representation of the layer.

8. The method for recommending potential suppliers based on a knowledge graph-enhanced relational graph convolutional network according to claim 1, characterized in that: The specific steps of optimizing and training the knowledge graph enhanced relational graph convolutional network model are as follows: Using a set of labeled supplier-customer interaction data, including positive and negative samples, that is, samples of real cooperation and samples of non-real cooperation, the learning goal is to minimize the difference between the predicted score and the true label through the binary cross-entropy loss function. After training, the model outputs a probability score to quantify the possibility of cooperation between each pair of supplier-customer. By ranking these scores, the system generates a personalized supplier recommendation list, where suppliers with higher rankings have a greater chance of becoming suitable partners.

9. The method for recommending potential suppliers based on a knowledge graph-enhanced relational graph convolutional network according to claim 1, characterized in that: The specific method of predicting the results is as follows: ； in, Represents the prediction results, represents the activation function, Represents the customer entity The relational graph convolutional network model enhanced by knowledge graph The final feature embedding obtained after layer processing, Represents a supplier entity The relational graph convolutional network model enhanced by knowledge graph The final feature embedding obtained after layer processing, Represents a learnable weight matrix for subsequent model training.

10. The method for recommending potential suppliers based on a knowledge graph-enhanced relational graph convolutional network according to claim 1, characterized in that: The specific method of evaluating the prediction results is as follows: Four indicators are used to evaluate the model, namely the area under the receiver operating characteristic curve, precision, overall accuracy and recall. Sample k and sample set for evaluation are set. Sample k belongs to the sample set, and the sample set includes 20, 50 and 100.

Images