Supplier intelligent recommendation method and recommendation system for automobile part purchase

Abstract

The invention belongs to the technical field of supplier recommendation, and particularly relates to an intelligent supplier recommendation method and system for automobile part purchase. The recommendation method comprises the following steps: constructing a knowledge graph composed of nodes and edges according to a historical order of a purchaser; according to a purchase demand, determining an accessory model node or an adaptive vehicle model node as a demand node, and based on a Monte Carlo Dropout mechanism, screening out a plurality of qualified supplier nodes of the demand node from the knowledge graph by the graph neural network model; based on a self-adaptive multi-objective optimization function, screening out a plurality of potential supplier nodes from all qualified supplier nodes; and calculating the side attention score of each potential supplier node, and outputting the first k suppliers of which the side attention scores are arranged in a descending order as recommendation degrees in a descending order. On the basis of integrating multi-source heterogeneous data, scientific and reasonable supplier recommendation can be carried out in combination with purchase demands and preferences.

Description

Technical Field

[0001] This application belongs to the field of supplier recommendation technology, and in particular relates to an intelligent supplier recommendation method and system for automotive parts procurement. Background Technology

[0002] With the rapid popularization of automobiles, the demand for spare parts from vehicle manufacturers and the maintenance service market is constantly increasing.

[0003] In traditional procurement models, buyers often rely on personal experience, fixed supplier lists, or simple price comparisons to select suppliers.

[0004] This process is not only time-consuming and labor-intensive, but also heavily influenced by the buyer's personal experience and preferences, resulting in highly subjective choices and recommendations that are biased towards a single dimension. This can lead to the selection of suppliers with irreparable shortcomings in certain areas. Furthermore, supplier data is scattered across multiple sources, including order systems, quality inspection systems, and financial systems. Data silos exist across different aspects of the same supplier, making it difficult for the buyer to establish effective connections between these multi-dimensional data and to comprehensively assess the supplier's overall capabilities.

[0005] Therefore, how to make scientific and reasonable supplier recommendations based on the integration of multi-source heterogeneous data and the combination of procurement needs and preferences has become an urgent problem to be solved in this field. Summary of the Invention

[0006] The purpose of this application is to overcome the shortcomings of the prior art and provide a supplier intelligent recommendation method for automotive parts procurement, which can make scientific and reasonable supplier recommendations based on the integration of multi-source heterogeneous data and the combination of procurement needs and preferences.

[0007] To achieve the above objectives, this application adopts the following technical solution: A supplier intelligent recommendation method for automotive parts procurement includes the following steps: S1. Based on the historical orders of the purchasers, construct a knowledge graph G consisting of nodes and edges; the nodes include supplier nodes, part model nodes, compatible vehicle model nodes, and purchaser nodes; S2, based on the procurement needs, determine the part model node or the compatible vehicle model node as the demand node. The graph neural network model GNN is based on the Monte Carlo Dropout mechanism to select several qualified supplier nodes for the demand node in the knowledge graph G. S3, based on an adaptive multi-objective optimization function, further selects several potential supplier nodes from all qualified supplier nodes; S4. Calculate the edge attention score of each potential supplier node, and output the suppliers corresponding to the top k potential supplier nodes with the edge attention scores in descending order as the recommendation level.

[0008] Preferably, in S1, the following also includes:

[0009] S11: The buyer tracks historical orders and obtains relevant data on auto parts; the relevant data includes order number, part model, compatible vehicle model, supplier, buyer, part price, delivery cycle, and supplier reputation; S12 uses the part model, compatible vehicle model, supplier, and purchaser as nodes. By connecting two nodes that appear in the same historical order with an edge, a basic graph BS is formed. Each edge contains a relationship type and corresponding relationship content between the two nodes in the same historical order. S13, after hierarchical aggregation and node embedding of the basic graph BS based on the graph neural network model GNN, the knowledge graph G is obtained; the hierarchical aggregation and node embedding is performed using the following formula: ; in, Indicates that node v is at the th Layer embedding vectors, =1,...; Represents a non-linear activation function; This represents the embedding vector of neighbor node u at layer l; This is the relation weight matrix; This represents the set of neighboring nodes that are adjacent to node v under relation type r; Represents aggregate functions under relation type r; This represents all relationship types in the base graph (BS), including price relationship types, fit relationship types, reputation relationship types, and delivery cycle relationship types; the embedding vector of each node in the first layer is obtained by initializing the graph neural network model (GNN) based on the node information.

[0010] Preferably, in S2, the following also includes:

[0011] S21, ensure that the Monte Carlo Dropout value of the graph neural network model GNN is within the set range of enabled values; S22. Based on the procurement requirements, the accessory model node or the compatible vehicle model node is determined as the demand node. The graph neural network model GNN performs N forward propagation in the knowledge graph G based on the demand node to obtain N preliminary qualified supplier nodes. S23, using the mean vector of N initially qualified supplier nodes. and variance As a filtering criterion, after confidence filtering and deduplication of the N initially qualified supplier nodes, the final M supplier nodes are obtained as qualified supplier nodes for the demand node: ; ; in, This indicates that the z-th preliminarily qualified supplier node is in the knowledge graph G. Layer embedding vector; It represents the square of the modulus.

[0012] Preferably, the training of the graph neural network model GNN to obtain preliminary qualified supplier nodes through forward propagation also includes the following:

[0013] Step 1a: Obtain a batch of historical orders and construct a knowledge graph PG for training; Step 2a: Using one of the parts model, compatible vehicle model, and supplier node as the demand node, the pairing relationship of the parts model, compatible vehicle model, and supplier that actually exist in the same historical order and contain the demand node is marked as a positive sample. All other nodes and relationships between nodes are marked as negative samples. This forms the subgraph training data centered on the current demand node.

[0014] Step 3a: The training data of the subgraph centered on a certain demand node is fed into the graph neural network model GNN to be trained. After normalizing and vectorizing each node and the relationship between nodes in the batch training data, the graph neural network model GNN learns the potential connections and differences between positive and negative samples in the batch training data of the demand node.

[0015] Preferably, the following is included after step 3a:

[0016] Step 4a: After constructing the knowledge graph VG for verification, select several demand nodes from it, use the demand nodes as samples of the current verification batch, and use the actual part models, compatible vehicle models, and supplier pairing relationships of these demand nodes as the real labels of the samples corresponding to the current verification batch. Step 5a: Using the validation knowledge graph VG, the current validation batch samples, and the true labels of the validation batch, the training effect of the graph neural network model GNN is validated and iteratively optimized: After calculating the joint loss function L of the current validation batch, the model parameters of the graph neural network model GNN are iteratively updated in the direction of decreasing gradient of the joint loss function of the current validation batch. ; in, and These represent the local loss weights and global loss weights of the training data in the current validation batch b, respectively. ; and These represent the local loss function and the global loss function of the current validation batch, respectively.

[0017] Preferably, step 5a further includes the following: The graph neural network model GNN outputs E prediction results based on the A samples of the current verification batch and the verification knowledge graph VG. Each prediction result contains a set of accessory models, compatible vehicle models, and supplier pairing relationships. The expressions for the local loss function and global loss function of the current validation batch are as follows: ; ; in, This represents the embedding vector of the accessory model node in the x-th prediction result. This represents the embedding vector of the matching vehicle model node in the x-th prediction result; This represents the veracity value of the x-th prediction result. If the x-th prediction result exists in the true labels of the current verification batch, then... ,otherwise ; express function; This represents the distribution of the true neighbor nodes of the g-th sample in the current validation batch; This represents the predicted neighbor node distribution of the g-th sample in the current validation batch; express Compared to The KL divergence.

[0018] Preferably, S3 also includes the following: S31, Based on several sub-objective functions of the supplier node, construct an adaptive multi-objective optimization function F: ; in, This represents the sub-objective function of total cost corresponding to selecting a certain supplier to meet the procurement requirements; This represents the sub-objective function for the risk of delays in choosing a particular supplier to meet procurement requirements; This represents the reputation sub-objective function corresponding to selecting a certain supplier to meet the procurement requirements; This indicates the probability that the parts provided by a particular supplier meet industry standards and satisfy the procurement requirements. Standard deviation = ; Indicates business constraints; , , , These represent the first, second, third, and fourth preference weights in the current recommendation round t, respectively. and , , , All values ​​are positive integers; a recommendation round refers to the period from obtaining the current procurement needs to outputting the corresponding recommended suppliers; S32, among M qualified supplier nodes, obtain the Pareto optimal solution set when the adaptive multi-objective optimization function F reaches its maximum value; S33, calculate the adaptive multi-objective optimization function value corresponding to each Pareto solution in the Pareto optimal solution set, and take the supplier nodes corresponding to the first d Pareto solutions in descending order of the adaptive multi-objective optimization function values ​​as potential supplier nodes.

[0019] Preferably, updating the preference weights for the next recommendation round also includes the following: Step 1b: The purchasing personnel determine the first, second, third, and fourth preference configurations for the next recommendation round (t+1). , , , ;in, , , , All are positive numbers and ; Step 2b: Calculate the preference weight for the next recommendation round based on the preference weight of the current recommendation round and the preference configuration for the next recommendation round. ; ; ; ; in, This represents the second fusion coefficient. ; Updating the global loss weights for the next validation batch also includes the following:

[0020] Step 1c: Technicians determine the global loss preference configuration for the next validation batch (b+1). ;in, ; Step 2c: Calculate the global loss weight for the next validation batch based on the global loss weight of the current validation batch and the global loss preference configuration for the next validation batch. ; in, Represents the first fusion coefficient. .

[0021] Preferably, S4 also includes the following: ; ; in, This represents the relationship between buyer node B and the c-th potential supplier node. Edge attention score between them; Represents the node importance weight coefficient This represents the path from buyer node B to the c-th potential supplier node. Edge attention weights; express function; Represents the function of a multilayer perceptron; This represents the embedding vector of the buyer node B; Indicates potential supplier nodes Node embedding vector This represents the relationship between buyer node B and the c-th potential supplier node. The relationship type of the edges connecting them; This represents the relationship between buyer node B and the c-th potential supplier node. The encoding vector of the relationship type of the edges between them; This represents the relationship between buyer node B and the c-th potential supplier node. The total number of edges contained in all paths between them; Represents the path attenuation coefficient. .

[0022] This application also provides a supplier intelligent recommendation system for automotive parts procurement, comprising: a knowledge graph construction module, a first screening module, a second screening module, and a third screening module; the knowledge graph construction module is used to construct a knowledge graph based on the buyer's historical orders and then send it to the first screening module; the first screening module is used to screen out several qualified supplier nodes based on procurement needs and then send them to the second screening module; the second screening module is used to further screen out several potential supplier nodes from all qualified supplier nodes based on an adaptive multi-objective optimization function and then send them to the third screening module; the third screening module is used to calculate the edge attention score of each potential supplier node, and output the top k potential supplier nodes in descending order of edge attention score as suppliers ranked in descending order of recommendation degree; each module is programmed or configured to execute the steps of the above-described supplier intelligent recommendation method for automotive parts procurement.

[0023] The beneficial effects of this application are as follows: (1) The intelligent supplier recommendation method for automotive parts procurement proposed in this application can make scientific and reasonable supplier recommendations based on the integration of multi-source heterogeneous data and the combination of procurement needs and preferences.

[0024] (2) This application proposes a supplier intelligent recommendation method for automotive parts procurement. By constructing a knowledge graph based on historical orders, the heterogeneous data from multiple sources, such as parts models, compatible vehicle models, purchasers, and suppliers, are uniformly displayed in the knowledge graph in the form of nodes and edges, thus avoiding the problem of information silos.

[0025] (3) Apply for a supplier intelligent recommendation method for automotive parts procurement. Among the suppliers included in historical orders, the method is screened three times based on demand nodes. While comprehensively considering personalized needs, it also considers the comprehensive performance of each supplier in other aspects, so that the recommendation results are true and reasonable, and have high flexibility and practicality. The recommended suppliers meet personalized needs and do not have obvious shortcomings, so the recommendation results are more objective. It greatly improves the work efficiency of procurement personnel and avoids the situation where the subjective preferences of procurement personnel completely affect the selection of suppliers.

[0026] (4) This application proposes a supplier intelligent recommendation method for automotive parts procurement. It uses a hierarchical aggregated graph neural network to construct a knowledge graph. The knowledge graph retains information on multi-dimensional relationships such as parts price, compatibility, reputation, and delivery cycle. It also integrates this information with node information such as parts model, compatible vehicle model, purchaser, and supplier. This enables the graph neural network model GNN, based on the Monte Carlo Dropout mechanism, to accurately capture the potential connection between procurement needs and suppliers in the knowledge graph G, and to select several qualified supplier nodes corresponding to the demand nodes, thereby improving the comprehensiveness and accuracy of the final recommendation results.

[0027] (5) The confidence filtering and the method for updating preference weights in this application effectively avoid overfitting caused by noisy nodes and the preferences of purchasing personnel, thus ensuring the stability and rationality of the recommendation process.

[0028] (6) In this application, the global loss weight update for each verification batch and the optimization graph neural network model GNN use a joint loss function composed of local loss function and global loss function, which greatly improves the optimization efficiency, makes the optimization process more scientific and stable, and improves the optimization effect.

[0029] (7) In this application, the new preference weight and global loss weight can be dynamically updated by integrating the previous preference weight and global loss weight according to the preferences input by the purchaser or the adjustment focus of the technical personnel, so that the supplier recommendation results can be updated adaptively, meet the personalized needs of different procurement scenarios, and stably recommend high-quality suppliers. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a supplier intelligent recommendation method for automotive parts procurement, as described in this application. Detailed Implementation

[0031] To make the technical solution of this application clearer and more explicit, the application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Solutions derived by those skilled in the art through equivalent substitution and conventional reasoning of the technical features of the technical solution of this application without creative effort all fall within the protection scope of this application.

[0032] like Figure 1 The diagram shown is a flowchart of a supplier intelligent recommendation method for automotive parts procurement according to this application, including the following steps: S1. Based on the historical orders of the purchasers, construct a knowledge graph G consisting of nodes and edges; the nodes include supplier nodes, part model nodes, compatible vehicle model nodes, and purchaser nodes; S2, based on the procurement needs, determine the part model node or the compatible vehicle model node as the demand node. The graph neural network model GNN is based on the Monte Carlo Dropout mechanism to select several qualified supplier nodes for the demand node in the knowledge graph G. S3, based on an adaptive multi-objective optimization function, further selects several potential supplier nodes from all qualified supplier nodes; S4. Calculate the edge attention score of each potential supplier node, and output the suppliers corresponding to the top k potential supplier nodes with the edge attention scores in descending order as the recommendation level.

[0033] S1 also includes the following sub-steps: S11 allows buyers to track historical orders and obtain relevant data on auto parts; this data includes order number, part model, compatible vehicle model, supplier, buyer, part price, delivery cycle, and supplier reputation.

[0034] By tracking a historical order, we can obtain: a supplier and a buyer, more than one part model, the compatible vehicle models corresponding to the buyer and the part model, the part price, the actual delivery cycle of a certain part model in the current historical order, and the supplier's reputation at the end of the current order.

[0035] The supplier's reputation at the end of the current order is scored by the buyer based on factors such as the estimated delivery time, the actual delivery time, and the percentage of qualified products delivered.

[0036] Technicians can obtain and track historical orders from several buyers over a period of time, based on their needs.

[0037] S12 uses the part model, compatible vehicle model, supplier, and purchaser as nodes. Connecting two nodes that appear in the same historical order with an edge forms the basic graph BS. Each edge contains a relationship type and corresponding relationship content between the two nodes in the same historical order. The relationship type and corresponding relationship content can be directly obtained from the relevant data in the corresponding historical order.

[0038] For example, if two nodes appear in m different historical orders, then there are at least m edges between these two nodes. If only one historical order contains nodes A and B, and that historical order contains the relationship content corresponding to the three relationship types between nodes A and B, then there are three edges between nodes A and B.

[0039] It should be noted that a historical order may contain a part model, but the buyer may have two compatible vehicle models for that part model. In this case, these two compatible vehicle models are treated as two separate nodes.

[0040] S13, after hierarchical aggregation and node embedding of the basic graph BS based on the graph neural network model GNN, the knowledge graph G is obtained; the hierarchical aggregation and node embedding is performed using the following formula: ; in, Indicates that node v is at the th Layer embedding vectors, =1,...; This represents a non-linear activation function, which in this embodiment is the LeakyReLU activation function; This represents the embedding vector of neighbor node u at layer l; This is the relation weight matrix; Represents the set of neighboring nodes that are adjacent to node v under relation type r; neighboring node u is only related to node v and relation type r, and is independent of the level; This represents the aggregate functions under relation type r, which are predefined by technical personnel. Aggregate functions include mean function, weighted sum function, max pooling function, attention aggregation function, etc. This represents all relationship types in the basic graph (BS), including: price relationship type, fit relationship type, reputation relationship type, delivery cycle relationship type, etc.; the embedding vector of each node in the first layer is obtained by the graph neural network model (GNN) after initialization and assignment based on the node information, and is a known vector.

[0041] S2 also includes the following: S21, ensure that the Monte Carlo Dropout value of the graph neural network model GNN is within the set range of enabled values.

[0042] In this embodiment, the range of the Monte Carlo Dropout value setting is 0.1 to 0.3.

[0043] S22, based on the procurement requirements, determine the part model node or the compatible vehicle model node as the demand node. The graph neural network model GNN performs N forward propagation operations in the knowledge graph G based on the demand node to obtain N preliminary qualified supplier nodes.

[0044] In this embodiment, N is an integer in the range of 20 to 50.

[0045] S23, using the mean vector of N initially qualified supplier nodes. and variance As a filtering criterion, after confidence filtering and deduplication of the N initially qualified supplier nodes, the final M supplier nodes are obtained as qualified supplier nodes for the demand node: ; ; in, This indicates that the z-th preliminarily qualified supplier node is in the knowledge graph G. Layer embedding vector; It represents the square of the modulus.

[0046] It's important to clarify here that duplicates may exist among the N initially qualified supplier nodes obtained in S22. Missing or abnormal historical orders can lead to some nodes being anomalous or uncertain. The confidence filtering in S23 aims to eliminate these anomalous and noisy nodes, ensuring that the qualified supplier nodes do not include noisy nodes. This further improves the robustness of subsequent calculations and the accuracy of the recommendation results, reducing uncertainty in the recommendation process.

[0047] Optionally, the training of the graph neural network model GNN to obtain preliminary qualified supplier nodes through forward propagation also includes the following: Step 1a: Obtain a batch of historical orders and construct a training knowledge graph PG. Step 2a: Using one of the parts model, compatible vehicle model, and supplier node as the demand node, the pairing relationship of the parts model, compatible vehicle model, and supplier that actually exist in the same historical order and contain the demand node is marked as a positive sample. All other nodes and relationships between nodes are marked as negative samples. This forms the subgraph training data centered on the current demand node.

[0048] Step 3a: The training data of the subgraph centered on a certain demand node is fed into the graph neural network model GNN to be trained. After normalizing and vectorizing each node and the relationship between nodes in the batch training data, the graph neural network model GNN learns the potential connections and differences between positive and negative samples in the batch training data of the demand node.

[0049] Optionally, the following may also be included after step 3a: Step 4a: After constructing the knowledge graph VG for verification, select several requirement nodes from it, use the requirement nodes as samples of the current verification batch, and use the actual part models, compatible vehicle models, and supplier pairing relationships of these requirement nodes as the real labels of the samples corresponding to the current verification batch.

[0050] Step 5a: Using the validation knowledge graph VG, the current validation batch samples, and the true labels of the validation batch, the training effect of the graph neural network model GNN is validated and iteratively optimized: After calculating the joint loss function L of the current validation batch, the model parameters of the graph neural network model GNN are iteratively updated in the direction of decreasing gradient of the joint loss function of the current validation batch. ; in, and These represent the local loss weights and global loss weights of the training data in the current validation batch b, respectively. ; and These represent the local loss function and the global loss function of the current validation batch, respectively.

[0051] Optionally, step 5a may also include the following: The graph neural network model GNN outputs E prediction results based on the A samples of the current verification batch and the verification knowledge graph VG. Each prediction result contains a set of accessory models, compatible vehicle models, and supplier pairing relationships. The expressions for the local loss function and the global loss function are as follows: ; ; in, This represents the embedding vector of the accessory model node in the x-th prediction result. This represents the embedding vector of the matching vehicle model node in the x-th prediction result. and These are all embedding vectors from the last layer of the graph neural network; This represents the veracity value of the x-th prediction result. If the x-th prediction result exists in the true labels of the current verification batch, then... ,otherwise ; express function; This represents the distribution of the true neighbor nodes of the g-th sample in the current verification batch. It can be obtained based on the verification knowledge graph VG, which is an existing technology and will not be elaborated here. This represents the predicted neighbor node distribution of the g-th sample in the current validation batch. It can be obtained based on the prediction results and the softmax function, which is existing technology and will not be elaborated here. express Compared to The KL divergence.

[0052] The graph neural network model (GNN) outputs the part model, compatible vehicle model, and supplier pairing relationship based on the demand node and the knowledge graph G. The recommended supplier output at the end must be a supplier in the pairing relationship output here.

[0053] It's important to emphasize here that during the training and iterative optimization of the Graph Neural Network (GNN) model, demand nodes can be supplier nodes. This is because the GNN's role is to learn the potential connections between pairs based on demand nodes in the training knowledge graph (PG) or to predict pairs based on demand nodes in the validation knowledge graph (VG). However, in practical applications, supplier nodes are output as the final recommendation result, so they cannot be demand nodes. Therefore, the term "demand nodes" mentioned in S22 does not include supplier nodes.

[0054] This application uses a joint loss function, consisting of local and global loss functions, to iteratively optimize the graph neural network model (GNN). This allows the GNN to more accurately output the pairing relationships corresponding to the demand nodes, thus improving the accuracy of the final recommended suppliers. The global loss function is used to improve the structural consistency between the predicted neighbor nodes and the actual neighbor nodes of the demand node. At the same time, the local loss function is used to improve the authenticity and practicality of the predicted pairing relationships (for example, if the purchasing department needs to purchase engines for a vehicle of model ρ, and the available engine models are ξ and τ, and supplier 1 does not produce engine model ξ; even if engine models ξ and τ exist in the knowledge graph G, and supplier 1 also exists, after iterative optimization, the graph neural network model GNN will basically not output pairing relationships such as "engine model ξ - vehicle model ρ - supplier 1" which lack practicality and authenticity).

[0055] S3 also includes the following: S31, based on several sub-objective functions of the supplier node, construct an adaptive multi-objective optimization function F.

[0056] S32, among M qualified supplier nodes, obtain the Pareto optimal solution set when the adaptive multi-objective optimization function F reaches its maximum value; S33, calculate the adaptive multi-objective optimization function value corresponding to each Pareto solution in the Pareto optimal solution set, and take the supplier nodes corresponding to the first d Pareto solutions in descending order of the adaptive multi-objective optimization function values ​​as potential supplier nodes.

[0057] In S31: ; in, This represents the total cost sub-objective function corresponding to selecting a certain supplier to meet the procurement requirements. The total cost includes price, freight, tariffs, etc. This represents the sub-objective function of the delay risk corresponding to selecting a certain supplier to meet the procurement requirements. The delivery delay risk of the same supplier is different for different models of parts that it can supply. This represents the reputation sub-objective function corresponding to selecting a certain supplier to meet the procurement requirements. The delivery speed of the same supplier varies for different models of parts that it can supply. This indicates the probability that the parts provided by a particular supplier meet industry standards and satisfy the procurement requirements. Standard deviation = ; These can represent business constraints, such as capacity constraints, blacklist / whitelist constraints, and upper limit threshold constraints for single dependency. , , , These represent the first, second, third, and fourth preference weights in the current recommendation round t, respectively. and , , , All values ​​are positive integers; a recommendation round refers to the period from obtaining the current procurement needs to outputting the corresponding recommended suppliers.

[0058] Sub-objective function based on supplier nodes ~ The specific calculation formula is existing technology and will not be elaborated here. The first, second, third, and fourth preference weights for each recommendation round can be manually set by the procurement personnel based on their procurement needs and preferences.

[0059] It should be noted here that, in order to more effectively screen potential supplier nodes from qualified supplier nodes, a multi-faceted objective (i.e., a sub-objective function) is used. ~ A multi-objective optimization function. Although we hope the optimal solution can make the sub-objective function... ~ While it's possible to achieve maximum and minimum values ​​for all objectives, this is clearly impossible, as improvements in some objectives would come at the expense of others. Furthermore, when using Monte Carlo Dropout for uncertainty estimation, we might simultaneously consider multiple objectives (such as accuracy and uncertainty). Therefore, here we solve for the Pareto set to balance the various sub-objective functions. The potential supplier nodes identified by the Pareto set will not exhibit poor performance in any one aspect, but rather demonstrate good performance across all aspects.

[0060] Optionally, updating the preference weights for the next recommendation round also includes the following: Step 1b: The purchasing personnel determine the first, second, third, and fourth preference configurations for the next recommendation round (t+1). , , , ;in, , , , All are positive numbers and .

[0061] Procurement personnel can assign values ​​to different preference configurations based on their own procurement preferences or the procurement needs of the next recommendation round. For example, if they want to increase the proportion of lower-priced component suppliers in the next recommendation round, they can increase the proportion of lower-priced component suppliers in that round. The value.

[0062] Step 2b: Calculate the preference weight for the next recommendation round based on the preference weight of the current recommendation round and the preference configuration for the next recommendation round. ; ; ; ; in, This represents the second fusion coefficient, which is pre-assigned by the technical personnel. .

[0063] Optionally, updating the global loss weights for the next validation batch also includes the following: Step 1c: Technicians determine the global loss preference configuration for the next validation batch (b+1). ;in, .

[0064] Step 2c: Calculate the global loss weight for the next validation batch based on the global loss weight of the current validation batch and the global loss preference configuration for the next validation batch. ; in, This represents the first fusion coefficient, which is pre-assigned by the technical personnel. .

[0065] It's important to note that in this application, neither the procurement personnel nor the technical personnel directly adjust the weights when updating the preference weights for the next recommendation round nor the global loss weights for the next validation batch. Instead, they directly change the preference configuration for the next round / batch. Then, based on this preference configuration, the corresponding weights for the next round / batch are calculated by fusing it with the weights of the current round. This method ensures that the weights for the next round / batch reflect individual needs without abrupt changes compared to the previous round / batch. This allows for smoother and more stable weight updates for each round / batch, closely following the preference configuration trend and reducing noise introduced by weight abrupt changes. It also prevents overfitting of the Graph Neural Network (GNN) model during data processing or adaptive multi-objective optimization function solving.

[0066] If procurement personnel consistently maintain a high value for a certain preference configuration, after several rounds of weight updates, the weights will gradually synchronize with the trend of the preference configuration values. This also indicates that during the supplier recommendation process, procurement personnel do indeed have a certain specific procurement preference (such as a preference for suppliers with lower costs) over a long period of time. In this case, the specific values ​​of the updated preference weights will not lead to overfitting.

[0067] By combining the solution of the Pareto set of the multi-objective optimization function with the indirect personalized adjustment of the preference weights in the multi-objective optimization function, the resulting potential supplier nodes not only perform well in all aspects (without any obvious shortcomings), but also ensure better performance in aspects corresponding to personalization (such as lower supplier costs or better quality auto parts), making the final recommendation results more robust and scientific.

[0068] S4 also includes the following: ; ; in, This represents the relationship between buyer node B and the c-th potential supplier node. Edge attention score between them; The weight coefficient representing the importance of nodes is a known quantity that can be set manually or calculated using a graph neural network. This is the relation weight matrix; This represents the path from buyer node B to the c-th potential supplier node. Edge attention weights; express function; Represents the function of a multilayer perceptron; This represents the embedding vector of the buyer node B; Indicates potential supplier nodes The node's embedding vector; and These are all embedding vectors from the last layer of the graph neural network; This represents the relationship between buyer node B and the c-th potential supplier node. The relationship type of the edges connecting them; This represents the relationship between buyer node B and the c-th potential supplier node. The encoding vector of the relationship type of the edges between them; This represents the relationship between buyer node B and the c-th potential supplier node. The total length of all paths between them, that is, the total number of edges contained in all paths, which is also the total number of hops; Represents the path attenuation coefficient. It is pre-set by technicians.

[0069] A higher edge attention score between a potential supplier node and a buyer node indicates a higher degree of authenticity and reasonableness in the relationship between the potential supplier and the buyer. The edge attention score also considers the total path length between the buyer and potential supplier nodes. This addresses situations where "some potential suppliers can supply goods but have never cooperated with the buyer." In such cases, the total path length between the buyer and potential supplier nodes is often large, leading to a lower edge attention score. This might be because the potential supplier has an exclusive supply agreement with other buyers, so even if the potential supplier is recommended, it won't appear in the top recommendation rankings.

[0070] This application presents a supplier intelligent recommendation method for automotive parts procurement. Although the initial starting point is the buyer's historical orders, it does not only recommend suppliers that the buyer has previously cooperated with. This is because the recommendation is based on a graph neural network model (GNN) that outputs the part model, the compatible vehicle model, and the supplier pairing relationship based on the demand node and the knowledge graph (G). Therefore, even if buyer A has never cooperated with supplier #1, but supplier #1 has supplied goods to other buyers and the part model provided by supplier #1 is compatible with buyer A's vehicle model, then supplier #1 may appear in the list of recommendations to buyer A.

[0071] This application presents an intelligent supplier recommendation method for automotive parts procurement. From suppliers included in historical orders, a triple screening process is performed based on demand nodes. This method comprehensively considers personalized needs (i.e., procurement personnel preferences) as well as the overall performance of each supplier in other aspects, ensuring that the recommendation results are accurate, reasonable, highly flexible, and practical. The recommended suppliers meet personalized needs without significant weaknesses, resulting in more objective recommendations. This significantly improves the work efficiency of procurement personnel and avoids situations where subjective preferences completely influence supplier selection.

[0072] This application presents a supplier intelligent recommendation method for automotive parts procurement. By constructing a knowledge graph based on historical orders, it uniformly displays heterogeneous data from multiple sources, such as parts models, compatible vehicle models, purchasers, and suppliers, in the form of nodes and edges, thus avoiding the problem of information silos.

[0073] This application presents an intelligent supplier recommendation method for automotive parts procurement. It employs a hierarchical aggregated graph neural network to construct a knowledge graph. This knowledge graph retains information on multi-dimensional relationships such as parts price, compatibility, reputation, and delivery cycle. Furthermore, it integrates this multi-dimensional relationship information with node information such as parts model, compatible vehicle type, purchaser, and supplier. This allows the graph neural network model (GNN) based on the Monte Carlo Dropout mechanism to accurately capture the potential connections between procurement needs and suppliers within the knowledge graph G, filtering out several qualified supplier nodes corresponding to the demand nodes, thus improving the comprehensiveness and accuracy of the final recommendation results.

[0074] The confidence filtering and preference weight update methods in this application effectively avoid overfitting caused by noisy nodes and purchasing personnel preferences (avoiding excessive sensitivity of the recommendation results to noisy nodes or purchasing personnel preferences), ensuring the stability and rationality of the recommendation process.

[0075] This application uses a joint loss function composed of local and global loss functions for updating the global loss weights for each validation batch and optimizing the graph neural network model GNN. This significantly improves optimization efficiency, makes the optimization process more scientific and stable, and enhances the optimization effect.

[0076] In this application, the system can dynamically update the preference weights and global loss weights based on the preferences input by the purchaser (such as "price priority" or "quality priority") or the shift in the focus of technical personnel. This is achieved by integrating the previous preference weights and global loss weights, enabling the supplier recommendation results to be updated adaptively. This satisfies the personalized needs of different procurement scenarios while stably recommending high-quality suppliers. (For example, if a supplier performs well in all aspects, it will still appear in the output recommendation list even if the purchaser prefers "price priority"; if the purchaser prefers "quality priority" in the next round of procurement, the supplier will still appear in the output recommendation list, but with a different ranking.)

[0077] This application also provides a supplier intelligent recommendation system for automotive parts procurement, including: The knowledge graph construction module, the first filtering module, the second filtering module, and the third filtering module; The knowledge graph construction module is used to build a knowledge graph based on the buyer's historical orders and then send it to the first filtering module; The first screening module is used to select several qualified supplier nodes based on procurement needs and then send them to the second screening module. The second screening module is used to further screen out several potential supplier nodes from all qualified supplier nodes based on an adaptive multi-objective optimization function, and then send them to the third screening module. The third filtering module is used to calculate the edge attention score of each potential supplier node, and output the top k potential supplier nodes in descending order of edge attention score as suppliers ranked in descending order of recommendation degree.

[0078] Each module is programmed or configured to execute the steps of the above-described intelligent supplier recommendation method for automotive parts procurement.

[0079] The technologies, shapes, and structures not described in detail in this application are all well-known technologies. It should also be noted that the above are merely preferred embodiments of this application and are not intended to limit the scope of this application. The components or steps in the embodiments of this application can be decomposed and / or recombined, and these decompositions and / or recombinations should be considered as equivalent solutions of this application and should all fall within the protection scope of this application.

Claims

1. A supplier intelligent recommendation method for automotive parts procurement, characterized in that, Includes the following steps: S1. Based on the historical orders of the purchasers, construct a knowledge graph G consisting of nodes and edges; the nodes include supplier nodes, part model nodes, compatible vehicle model nodes, and purchaser nodes; S2, based on the procurement needs, determine the part model node or the compatible vehicle model node as the demand node. The graph neural network model GNN is based on the Monte Carlo Dropout mechanism to select several qualified supplier nodes for the demand node in the knowledge graph G. S3, based on an adaptive multi-objective optimization function, further selects several potential supplier nodes from all qualified supplier nodes; S4. Calculate the edge attention score of each potential supplier node, and output the suppliers corresponding to the top k potential supplier nodes with the edge attention scores in descending order as the recommendation level.

2. The intelligent supplier recommendation method for automotive parts procurement according to claim 1, characterized in that, S1 also includes the following: S11: The buyer tracks historical orders and obtains relevant data on auto parts; the relevant data includes order number, part model, compatible vehicle model, supplier, buyer, part price, delivery cycle, and supplier reputation; S12 uses the part model, compatible vehicle model, supplier, and purchaser as nodes. By connecting two nodes that appear in the same historical order with an edge, a basic graph BS is formed. Each edge contains a relationship type and corresponding relationship content between the two nodes in the same historical order. S13, after hierarchical aggregation and node embedding of the basic graph BS based on the graph neural network model GNN, the knowledge graph G is obtained; the hierarchical aggregation and node embedding is performed using the following formula: ； in, Indicates that node v is at the th Layer embedding vectors, =1,...; Represents a non-linear activation function; This represents the embedding vector of neighbor node u at layer l; This is the relation weight matrix; This represents the set of neighboring nodes that are adjacent to node v under relation type r; Represents aggregate functions under relation type r; It represents all relationship types in the base graph (BS), including price relationship types, fit relationship types, reputation relationship types, and delivery cycle relationship types; The embedding vector of each node in the first layer is obtained by initializing the graph neural network model GNN based on the node information.

3. The intelligent supplier recommendation method for automotive parts procurement according to claim 1, characterized in that, S2 also includes the following: S21, ensure that the Monte Carlo Dropout value of the graph neural network model GNN is within the set range of enabled values; S22. Based on the procurement requirements, the accessory model node or the compatible vehicle model node is determined as the demand node. The graph neural network model GNN performs N forward propagation in the knowledge graph G based on the demand node to obtain N preliminary qualified supplier nodes. S23, using the mean vector of N initially qualified supplier nodes. and variance As a filtering criterion, after confidence filtering and deduplication of the N initially qualified supplier nodes, the final M supplier nodes are obtained as qualified supplier nodes for the demand node: ； ； in, This indicates that the z-th preliminarily qualified supplier node is in the knowledge graph G. Layer embedding vector; It represents the square of the modulus.

4. The intelligent supplier recommendation method for automotive parts procurement according to claim 3, characterized in that, The training graph neural network model (GNN) obtains preliminary qualified supplier nodes through forward propagation, and also includes the following: Step 1a: Obtain a batch of historical orders and construct a knowledge graph PG for training; Step 2a: Take one of the parts model, compatible vehicle model, and supplier node as the demand node. Label the pairing relationship of the parts model, compatible vehicle model, and supplier that actually exist in the same historical order and contain the demand node as positive samples. Label the remaining nodes and relationships between nodes as negative samples. This forms the subgraph training data centered on the current demand node. Step 3a: The training data of the subgraph centered on a certain demand node is fed into the graph neural network model GNN to be trained. After normalizing and vectorizing each node and the relationship between nodes in the batch training data, the graph neural network model GNN learns the potential connections and differences between positive and negative samples in the batch training data of the demand node.

5. The intelligent supplier recommendation method for automotive parts procurement according to claim 4, characterized in that, The following is also included after step 3a: Step 4a: After constructing the knowledge graph VG for verification, select several demand nodes from it, use the demand nodes as samples of the current verification batch, and use the actual part models, compatible vehicle models, and supplier pairing relationships of these demand nodes as the real labels of the samples corresponding to the current verification batch. Step 5a: Using the validation knowledge graph VG, the current validation batch samples, and the true labels of the validation batch, the training effect of the graph neural network model GNN is validated and iteratively optimized: After calculating the joint loss function L of the current validation batch, the model parameters of the graph neural network model GNN are iteratively updated in the direction of decreasing gradient of the joint loss function of the current validation batch. ； in, and These represent the local loss weights and global loss weights of the training data in the current validation batch b, respectively. ; and These represent the local loss function and the global loss function of the current validation batch, respectively.

6. The intelligent supplier recommendation method for automotive parts procurement according to claim 5, characterized in that, Step 5a also includes the following: The graph neural network model GNN outputs E prediction results based on the A samples of the current verification batch and the verification knowledge graph VG. Each prediction result contains a set of accessory models, compatible vehicle models, and supplier pairing relationships. The expressions for the local loss function and global loss function of the current validation batch are as follows: ； ； in, This represents the embedding vector of the accessory model node in the x-th prediction result. This represents the embedding vector of the matching vehicle model node in the x-th prediction result; This represents the veracity value of the x-th prediction result. If the x-th prediction result exists in the true labels of the current verification batch, then... ,otherwise ; express function; This represents the distribution of the true neighbor nodes of the g-th sample in the current validation batch; This represents the predicted neighbor node distribution of the g-th sample in the current validation batch; express Compared to The KL divergence.

7. The intelligent supplier recommendation method for automotive parts procurement according to claim 3, characterized in that, S3 also includes the following: S31, Based on several sub-objective functions of the supplier node, construct an adaptive multi-objective optimization function F: ； in, This represents the sub-objective function of total cost corresponding to selecting a certain supplier to meet the procurement requirements; This represents the sub-objective function for the risk of delays in choosing a particular supplier to meet procurement requirements; This represents the reputation sub-objective function corresponding to selecting a certain supplier to meet the procurement requirements; This indicates the probability that the parts provided by a particular supplier meet industry standards and satisfy the procurement requirements. Standard deviation = ; Indicates business constraints; , , , These represent the first, second, third, and fourth preference weights in the current recommendation round t, respectively. and , , , All values ​​are positive integers; a recommendation round refers to the period from obtaining the current procurement needs to outputting the corresponding recommended suppliers; S32, among M qualified supplier nodes, obtain the Pareto optimal solution set when the adaptive multi-objective optimization function F reaches its maximum value; S33, calculate the adaptive multi-objective optimization function value corresponding to each Pareto solution in the Pareto optimal solution set, and take the supplier nodes corresponding to the first d Pareto solutions in descending order of the adaptive multi-objective optimization function values ​​as potential supplier nodes.

8. A supplier intelligent recommendation method for automotive parts procurement according to claim 5 or 7, characterized in that: Updating the preference weights for the next recommendation round also includes the following: Step 1b: The purchasing personnel determine the first, second, third, and fourth preference configurations for the next recommendation round (t+1). , , , ;in, , , , All are positive numbers and ; Step 2b: Calculate the preference weight for the next recommendation round based on the preference weight of the current recommendation round and the preference configuration for the next recommendation round. ； ； ； ； in, This represents the second fusion coefficient. ; Updating the global loss weights for the next validation batch also includes the following: Step 1c: Technicians determine the global loss preference configuration for the next validation batch (b+1). ;in, ; Step 2c: Calculate the global loss weight for the next validation batch based on the global loss weight of the current validation batch and the global loss preference configuration for the next validation batch. ； in, Represents the first fusion coefficient. .

9. A supplier intelligent recommendation method for automotive parts procurement according to claim 2, characterized in that, S4 also includes the following: ； ； in, This represents the relationship between buyer node B and the c-th potential supplier node. Edge attention score between them; This represents the weighting coefficient indicating the importance of a node. This represents the path from buyer node B to the c-th potential supplier node. Edge attention weights; express function; Represents the function of a multilayer perceptron; This represents the embedding vector of the buyer node B; Indicates potential supplier nodes The node's embedding vector; This represents the relationship between buyer node B and the c-th potential supplier node. The relationship type of the edges connecting them; This represents the relationship between buyer node B and the c-th potential supplier node. The encoding vector of the relationship type of the edges between them; This represents the relationship between buyer node B and the c-th potential supplier node. The total number of edges contained in all paths between them; Represents the path attenuation coefficient. .

10. A supplier intelligent recommendation system for automotive parts procurement, characterized in that, include: The system comprises a knowledge graph construction module, a first screening module, a second screening module, and a third screening module. The knowledge graph construction module constructs a knowledge graph based on the buyer's historical orders and then sends it to the first screening module. The first screening module selects several qualified supplier nodes based on the procurement requirements and then sends them to the second screening module. The second screening module, based on an adaptive multi-objective optimization function, further selects several potential supplier nodes from all qualified supplier nodes and then sends them to the third screening module. The third screening module calculates the edge attention score of each potential supplier node and outputs the top k potential supplier nodes in descending order of edge attention score, which are then used as suppliers ranked in descending order of recommendation level. Each module is programmed or configured to perform the steps of the intelligent supplier recommendation method for automotive parts procurement as described in any of claims 1-9.

Images