Purchase supplier intelligent recommendation method and system based on artificial intelligence

Abstract

The invention discloses a purchase supplier intelligent recommendation method and system based on artificial intelligence, and relates to the supply chain management cross technology field, and the method comprises the steps: obtaining purchase events and supplier historical participation data, constructing a sparse participation matrix and a knowledge graph, and generating a joint input vector; a quality membership vector is calculated based on the joint input vector, a risk membership vector is obtained by applying a VIDER method, a fuzzy interaction weight is calculated in combination with a Sugeno model, a prediction recommendation score is generated by using the interaction weight, and encryption model parameters are obtained in combination with a BPR loss function; receiving the encrypted and uploaded model parameters, decrypting the model parameters, obtaining the model parameters by adopting a FedAvg algorithm, and updating the model parameters through a broadcast updating mechanism; and generating a node embedding vector by using a GraphSAGE model, and calculating a comprehensive risk score by using updated model parameters. According to the method, the accuracy, interpretability and sustainability of supplier recommendation are improved.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, and in particular to an intelligent supplier recommendation method and system based on artificial intelligence. Background Technology

[0002] Supplier recommendation technology has seen rapid development in the field of supply chain management. With advancements in big data and artificial intelligence, machine learning-based recommendation algorithms are increasingly being introduced into procurement scenarios. For example, collaborative filtering and content-based recommendation methods generate personalized recommendations by analyzing historical interaction data between buyers and suppliers. The application of knowledge graph technology in supply chain management has also significantly enhanced the semantic understanding capabilities of recommendation systems. By constructing complex relationship networks between suppliers, procurement events, and materials, and combining external information (such as news and financial reports), richer contextual support is provided for recommendations. The rise of federated learning and privacy-preserving computing technologies has further promoted research into distributed procurement systems, making data privacy protection possible in multi-party collaboration. However, existing supplier recommendation technologies still have several shortcomings. Traditional recommendation methods have limited effectiveness when dealing with sparse data, especially when procurement events and supplier participation records are highly sparse. The accuracy and robustness of the recommendations are difficult to guarantee. Existing methods are insufficient in modeling dynamic risk factors. For example, the impact of external events such as the supplier's financial situation, supply chain disruptions, or geopolitical risks is often not fully considered, resulting in a lack of adaptability of the recommendations in dynamic environments. Recommendation models based on a single dimension (such as considering only quality or cost) cannot fully capture the multidimensional characteristics of suppliers (such as quality, risk, ESG compliance, etc.), thus limiting the comprehensiveness and fairness of the recommendations. Existing systems also have insufficient protection of data privacy, especially in multi-party collaboration scenarios, where the sharing of sensitive data may lead to privacy leakage risks. Although the application of knowledge graphs has improved semantic understanding capabilities, improvements are still needed in terms of the real-time nature of dynamic updates and node embedding, especially how to quickly integrate external events into the graph and reflect them in the recommendation results. These shortcomings make it difficult for existing technologies to meet the comprehensive requirements of high accuracy, dynamic adaptation, and privacy protection when facing complex and ever-changing procurement scenarios. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides an intelligent recommendation method and system for procurement suppliers based on artificial intelligence, which solves the problems of existing technologies in terms of dynamic risk modeling, privacy protection, and dynamic updating of knowledge graphs.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an intelligent recommendation method for procurement suppliers based on artificial intelligence, which includes: constructing a sparse participation matrix and a knowledge graph by acquiring procurement events and historical participation data of suppliers, and generating a joint input vector; Based on the joint input vector, the quality membership vector is calculated, the VADER method is applied to obtain the risk membership vector, and the Sugeno model is combined to calculate the fuzzy interaction weights. The interaction weights are used to generate the predicted recommendation score, and the parameters of the encrypted model are obtained by combining the BPR loss function. The system receives encrypted uploaded model parameters, decrypts them, uses the FedAvg algorithm to obtain the model parameters, updates the model parameters through a broadcast update mechanism, uses the GraphSAGE model to generate node embedding vectors, calculates the comprehensive risk score using the updated model parameters, and updates the knowledge graph by combining the node embedding vectors with the comprehensive risk score. The updated risk attributes are obtained by updating the knowledge graph. The risk attributes are combined with the predicted recommendation score to obtain the final recommendation score of the supplier. A supplier recommendation list is generated based on the recommendation score. The recommendation hit rate is calculated using the supplier recommendation list. The recommendation strategy is updated based on the recommendation hit rate to achieve closed-loop adaptation.

[0006] As a preferred embodiment of the AI-based intelligent supplier recommendation method described in this invention, the method involves: constructing a sparse participation matrix and a knowledge graph by acquiring procurement event and supplier historical participation data, and generating a joint input vector. This is achieved by collecting procurement event logs, supplier historical participation records, and external events to construct the sparse participation matrix. , which contains matrix elements Constructing an initial knowledge graph When matrix elements When =1, the purchaser ID is one-hot encoded, and the requirement text is converted into a semantic feature vector using TF-IDF. This is then concatenated with the supplier's identifier vector to generate a joint input vector. .

[0007] As a preferred embodiment of the AI-based intelligent supplier recommendation method described in this invention, the method involves: calculating a quality membership vector based on a joint input vector; applying the VADER method to obtain a risk membership vector; calculating fuzzy interaction weights using a Sugeno model; generating a predicted recommendation score using the interaction weights; obtaining encrypted model parameters using the BPR loss function; extracting supplier quality ratings from the buyer's database; and normalizing the original supplier quality ratings to obtain the final quality rating. Three triangular fuzzy sets are defined using fuzzy inference methods for each quality rating. Membership degree is obtained using trigonometric membership functions. Calculate the quality membership vector Text information is generated using collected external event data related to suppliers. For each text Calculate sentiment scores and generate risk values. Quality rating for each supplier and risk value The difference between the actual value and the predicted value is calculated to obtain the scalar residual. and scalar residual Substitute the values ​​into the three triangular fuzzy sets respectively, and output the membership values ​​in each fuzzy set, relative to the scalar residuals. Applying the Gaussian fuzzy membership function, the fuzzy membership vector is obtained. ; Calculate the interaction weights using the Sugeno fuzzy inference model. The PROMETHEE model is used to generate pairwise comparisons of preference strengths, resulting in a comprehensive preference index. Based on interaction weights and comprehensive preference index Combined with the FM framework, predict recommendation scores Based on predictive recommendation rating Calculate the loss value of the BPR loss function. Calculate the loss function For the gradient of each parameter, update the model parameters, and encrypt only the subset of parameters that need to be shared using the Paillier homomorphic encryption algorithm.

[0008] As a preferred embodiment of the intelligent supplier recommendation method based on artificial intelligence described in this invention, the following steps are taken: receiving encrypted uploaded model parameters, decrypting them, obtaining model parameters using the FedAvg algorithm, and updating the model parameters through a broadcast update mechanism, indicating that the central server updates the encrypted model parameters from K local nodes. Decryption is performed using a weighted average calculated based on the FedAvg algorithm to determine the global model parameters. The central server asynchronously broadcasts the updated global model parameters to all participating local nodes via the network, using message queues, peer-to-peer push, or streaming communication. This allows some nodes to receive the updates within an acceptable delay. Once each local node receives the model parameters... Replace the current local model parameters as the starting point for the next round of training, continue local optimization until the model performance converges, and output the updated global FM model parameters. .

[0009] As a preferred embodiment of the AI-based intelligent supplier recommendation method described in this invention, the following steps are taken: GraphSAGE model is used to generate node embedding vectors; the comprehensive risk score is calculated using updated model parameters; and the node embedding vectors, combined with the comprehensive risk score, update the knowledge graph based on the initial knowledge graph. Generate intermediate state graphs that have not yet been embedded and updated. Using the trained GraphSAGE model to analyze the intermediate state graph Perform node embedding updates to obtain the latest node embedding vector. Calculate the edge weights after fusion The overall risk score is calculated and propagated based on the sentiment score of external events extracted using the VADER sentiment analysis method. , the latest embedding vector This serves as a new feature vector for suppliers, while simultaneously incorporating the comprehensive risk score calculated from risk propagation. As node attributes, the new feature vectors and node attributes are batch-written into the corresponding supplier nodes to generate a new knowledge graph. .

[0010] As a preferred embodiment of the intelligent supplier recommendation method based on artificial intelligence described in this invention, the following steps are taken: obtaining updated risk attributes based on an updated knowledge graph; combining the risk attributes with a predicted recommendation score to obtain a final supplier recommendation score; and generating a supplier recommendation list based on the recommendation score, which is based on the updated knowledge graph. Extracting the latest node information, including obtaining the embedding vector using the GraphSAGE model inference. As a structured representation of the supplier, obtain the updated risk attributes. Fusion prediction recommendation score Calculate supplier recommendation scores by combining contextual information enhanced by graph embedding with dynamic risk scores. The top 10 items were selected to form the Top-10 recommendation list.

[0011] As a preferred embodiment of the intelligent supplier recommendation method based on artificial intelligence described in this invention, the following steps are taken: The method involves calculating the recommendation hit rate using a supplier recommendation list, updating the recommendation strategy based on the hit rate to achieve closed-loop adaptive recommendation. This involves comparing the generated Top-10 recommendation list with the actual selections made by the purchaser in the ERP system, calculating the recommendation hit rate, combining the purchaser's human rating of the recommendation results, constructing a reward signal, and inputting it into a Q-learning reinforcement learning model. The Q-learning reinforcement learning model uses the weighted sum of the hit rate and human rating as the reward, and optimizes the strategy through a time-difference update mechanism.

[0012] Secondly, the present invention provides an intelligent recommendation system for procurement suppliers based on artificial intelligence, including a data acquisition and preprocessing module, used to acquire supplier data to construct a sparse participation matrix and an initial knowledge graph, and generate a joint input vector; The fuzzy reasoning and rating prediction module is used to calculate quality and risk membership vectors, and combines the Sugeno model and the PROMETHEE method to generate fuzzy interaction weights and predict recommendation ratings. The federated learning and model update module receives encrypted model parameters, aggregates global parameters using the FedAvg algorithm, and achieves collaborative optimization through a broadcast update mechanism. The graph embedding and risk assessment module is used to calculate a comprehensive risk score by combining global model parameters and VADER sentiment analysis, and to update the node features and edge weights of the knowledge graph. The recommendation generation and adaptive optimization module is used to update the knowledge graph, extract risk attributes, integrate prediction scores to generate a supplier recommendation list, calculate the recommendation hit rate, and optimize the recommendation strategy through Q-learning.

[0013] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the intelligent recommendation method for procurement suppliers based on artificial intelligence as described in the first aspect of the present invention.

[0014] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent recommendation method for procurement suppliers based on artificial intelligence as described in the first aspect of the present invention.

[0015] The beneficial effects of this invention are as follows: By constructing a sparse participation matrix and knowledge graph, combined with fuzzy inference methods and VADER sentiment analysis, the invention accurately captures supplier quality and dynamic risk characteristics, effectively improving recommendation accuracy and robustness; by utilizing the GraphSAGE model and FedAvg algorithm, it achieves dynamic updating of the knowledge graph and privacy protection under multi-party collaboration, solving the problems of slow response to external events and data privacy leakage in traditional methods; by optimizing recommendation ranking through the BPR loss function and combining the PROMETHEE and CRITIC methods to generate a comprehensive preference index, it ensures the balance and fairness of recommendation results in multiple dimensions, significantly improving the accuracy, adaptability, and security of supplier recommendations, and fully meeting the needs of complex procurement scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an AI-based intelligent supplier recommendation method in Example 1.

[0018] Figure 2 This is a schematic diagram of the structure of an AI-based intelligent recommendation system for procurement suppliers in Example 1.

[0019] Figure 3 This is a flowchart of the knowledge graph update process in Example 1. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, referring to Figures 1 to 3 This is the first embodiment of the present invention, which provides an intelligent supplier recommendation method based on artificial intelligence, including the following steps: S1. By acquiring procurement events and historical supplier participation data, construct a sparse participation matrix and knowledge graph, and generate a joint input vector; Based on the joint input vector, the quality membership vector is calculated, the VADER method is applied to obtain the risk membership vector, and the Sugeno model is combined to calculate the fuzzy interaction weights. The interaction weights are used to generate the predicted recommendation score, and the parameters of the encrypted model are obtained by combining the BPR loss function. Specifically, by acquiring procurement events and supplier historical participation data, a sparse participation matrix and knowledge graph are constructed, and a joint input vector is generated. Procurement event logs (purchaser ID, time zone, requirement text) and supplier historical participation records (delivery, contract) are obtained from the SAP ERP system through API interface. External events (supply chain disruptions, ESG reports, geopolitical risks, financial status) are collected from public channels (news, APP platforms) using Scrapy crawler. Construct a sparse participation matrix based on procurement event logs and supplier history. The rows correspond to procurement events e (approximately 1000), the columns correspond to suppliers s (approximately 1000), and the matrix elements... This indicates whether supplier s participated in event e. If it participated, it is marked as 1; otherwise, it is marked as 0 (indicating that it did not participate or did not respond). Structured data on suppliers, procurement events, and contract relationships were extracted from the SAP ERP system. Suppliers, purchasers, and materials were defined as nodes, and contract relationships, delivery transactions, and hierarchical relationships as edges, forming a basic knowledge graph framework. External relationships were supplemented using Named Entity Recognition (NER) and relationship extraction techniques, incorporating collected external events. Supplier quality ratings and historical delivery performance attributes were then used as node features, and transaction frequency as edge weights. This data was imported into the Neo4j graph database and cleaned through consistency checks and deduplication, resulting in an initial knowledge graph containing approximately 1000 nodes and thousands of relationships. Where V represents the set of all nodes in the graph, including suppliers, purchasers, and material entities, and E represents the set of all edges in the graph, i.e. the relationships between nodes, such as contractual relationships, delivery transactions, and hierarchical relationships.

[0024] when When =1, the purchaser ID is one-hot encoded (dimension S≈1000), the time zone information is normalized to [0,1], and the requirement text is transformed into a semantic feature vector through TF-IDF (vocabulary dimension M=5000). It is then concatenated with the supplier's identifier vector 1s to generate a joint input vector. .

[0025] By acquiring procurement events and historical supplier participation data from the SAP ERP system and combining this with Scrapy web crawling to collect external events, a sparse participation matrix and knowledge graph are constructed. This effectively solves the problem of insufficient recommendation accuracy caused by data sparsity in traditional methods. Named entity recognition (NER) and relation extraction techniques are used to supplement external association edges, and supplier quality ratings and transaction frequencies are integrated into the Neo4j graph database to form a dynamic knowledge graph containing approximately 1,000 nodes and thousands of relationships. This enhances the semantic modeling capability for complex supply chain relationships. By converting purchaser IDs, time zone information, and demand text into joint input vectors, efficient fusion of multi-dimensional features is achieved. This overcomes the limitations of traditional methods in responding to dynamic external events, improves the accuracy and adaptability of the recommendation system, and demonstrates significant advantages over existing technologies.

[0026] Furthermore, based on the joint input vector, a quality membership vector is calculated, and the risk membership vector is obtained by applying the VADER method. Fuzzy interaction weights are then calculated using the Sugeno model. Predictive recommendation scores are generated using these interaction weights, and encrypted model parameters are obtained by combining the BPR loss function. Supplier quality ratings (subjective ratings 1-5, based on contract performance, on-time delivery rate, and product quality feedback; for example, supplier A with an on-time delivery rate of 90% and a complaint rate of <2% is rated 4) are extracted from the buyer's database via API. The original supplier quality ratings are normalized and mapped to the [0,1] interval (e.g., a rating of 4 is normalized to 0.75 (i.e., (4-1) / 4=0.75)) to obtain the quality rating. Using the Mamdani fuzzy inference method, three triangular fuzzy sets are defined: "low" (e.g., parameters [0,0,0.3]), "medium" (e.g., [0.2,0.5,0.8]), and "high" (e.g., [0.7,1,1]). These parameters are set by procurement experts according to industry standards (e.g., the ISO 9001 quality system). Each quality rating is then assigned a specific parameter. Membership degree is obtained using trigonometric membership functions. The formula is: , Where c is the center of the fuzzy set. Given the width, for example, the center height is 0.85 and the width is 0.3, calculate the membership degree, for example... ,get , ∈[0,1] represents the membership degree of the quality rating belonging to the "low" fuzzy set. ∈[0,1] represents the membership degree of the quality rating belonging to the "medium" fuzzy set. ∈[0,1] represents the membership degree of the quality rating belonging to the "high" fuzzy set; A quality membership vector is formed based on the membership values ​​in the three sets: "low", "medium", and "high". ,Will Associated with the supplier ID, stored in a local MongoDB database; Text information is generated using collected external event data related to suppliers. The VADER sentiment analysis method was applied to each text. Calculate sentiment scores, ranging from [-1, 1], where negative scores (such as "supplier B bankruptcy rumors") are mapped to high risk, and positive scores (such as "supplier A obtains green certification") are mapped to low risk, generating risk values. And normalized to [0,1], the formula is: , Where a is the normalization factor; Quality rating for each supplier and risk value Predictions are made using a linear regression model (such as the LinearRegression model in scikit-learn) to obtain predicted values. and Calculate the difference between the actual value and the predicted value, i.e., the scalar residual. and scalar residual Substitute the values ​​into the three triangular fuzzy sets respectively, and output the membership values ​​in each fuzzy set (i.e., ...). (Classifying as "low, medium, or high deviation"), for each preset Sugeno fuzzy rule. (For example, "if the residual is low, then the interaction weight is high"), the input variable scalar residual Perform a conjunction operation on the corresponding membership degrees (usually taking the minimum value) to obtain the overall activation strength. scalar residual Normalized to [0,1], for scalar residuals Using Gaussian fuzzy membership functions, three Gaussian fuzzy sets are defined: "Low" (mean 0.15, variance 0.1), "Medium" (mean 0.5, variance 0.15), and "High" (mean 0.85, variance 0.1). The parameters are set by procurement experts based on industry risk standards. The parameters of the three Gaussian fuzzy sets are then substituted into the Gaussian membership function. Calculate the residual The membership degree of each fuzzy set, i.e., the fuzzy membership degree vector. ,For example =0.4 The Gaussian membership function formula is: , Where m is the mean. Standard deviation; Using the Sugeno fuzzy inference model, rules are defined (e.g., "if the quality is high and the risk is low, then output a weight of 0.9"), and fuzzy interaction weights are calculated using a weighted linear combination. The formula is: , The PROMETHEE model is used to generate pairwise comparisons of preference strengths, and CRITIC (standard deviation and correlation coefficient) is used for objective weighting. These are then combined to generate a comprehensive preference index. The formula is:

[0027] in, and For multi-objective criteria (such as supplier cost, quality, risk, and ESG), CRITIC automatically determines the weights based on the dispersion (standard deviation) and interrelationship (correlation coefficient) of each criterion in historical data. A larger standard deviation indicates higher information content, while a smaller correlation coefficient indicates stronger independence, comprehensively reflecting the importance of the indicators. For weighting; Based on the performance data of all suppliers in terms of characteristics (such as quality, cost, risk, ESG), and utilizing existing supplier data... Generate Preference Index ; Incorporating fuzzy interaction weights and comprehensive preference index Combined with the FM (Factor Factorization) framework, the recommendation score is predicted. The formula is: , in, For bias, For linear weights, For feature embedding (dimension 32). To enhance the learning of dynamic weights, S represents static features, such as supplier quality and cost, and M represents dynamic features, such as risk and ESG. and For the first and Embedding vectors of each feature The inner product of two feature embeddings represents the strength of the nonlinear interaction. and For the first and The value of each feature (e.g., 0 or 1). This is a feature co-occurrence indicator, and is only valid when both values ​​are 1. From the matrix Extract the participation record for each event e, positive samples. satisfy An "event-supplier" combination of 1 indicates that the supplier actually participated in the procurement event; negative samples Random sampling from suppliers who have never participated (i.e.) =0), ensuring that it belongs to the same event e as the positive sample, forming a comparison sample pair ( ); Based on predicted recommendation ratings To achieve the goal, the BPR loss function is used to optimize the model (the model refers to the model that integrates fuzzy membership, comprehensive preference index and personalized recommendation model (i.e., the aforementioned)). The predictive model's ranking ability is improved, enhancing the accuracy and robustness of the recommendation results. The loss value of the BPR loss function is calculated. (Scalar value), representing the average ranking loss of the current model across all positive and negative sample pairs. The smaller the value, the more accurate the model's ranking of participating behaviors. The formula is: , in, and These are the model's prediction scores for positive and negative samples, respectively. Let q be the sigmoid function, which maps the difference to the interval [0,1] to enhance gradient stability. Let q represent an edge in the graph (i.e., a relationship, such as the association between "supplier-purchaser" or "supplier-material"), and let I be the set of all edges in the graph associated with the supplier. Let q be the set of positive samples associated with edge q. The loss function is calculated using stochastic gradient descent (SGD). Update the model parameters based on the gradients of each parameter. Only for the subset of parameters that need to be shared. The Paillier homomorphic encryption algorithm is used for encryption, the generated ciphertext is sent to the central server via gRPC communication, and the encrypted data participates in federated aggregation to ensure that the original parameters are not leaked.

[0028] By extracting supplier quality ratings from the buyer's database and generating quality membership vectors using the Mamdani fuzzy inference method, and using VADER sentiment analysis to process external event data to generate risk membership vectors, and combining the Sugeno fuzzy inference model to calculate interaction weights, this approach overcomes the shortcomings of traditional methods in assessing supplier quality and dynamic risks, improving the accuracy and robustness of recommendation results. It employs the PROMETHEE and CRITIC methods to generate a comprehensive preference index, integrating multi-dimensional features to address the limitations of single-dimensional recommendations and ensure the comprehensiveness and fairness of the results. A factorization machine (FM) framework is used to integrate fuzzy interaction weights and preference indices to generate predictive recommendation scores, and the BPR loss function is combined to optimize ranking capabilities, improving the accuracy and stability of the recommendation system. Paillier homomorphic encryption and federated aggregation mechanisms are used to achieve parameter sharing while protecting data privacy in multi-party collaborative scenarios, addressing the shortcomings of traditional methods in data sensitivity, improving the accuracy, adaptability, and fairness of supplier recommendations, and enhancing the system's privacy protection and dynamic response capabilities. This provides an efficient and reliable solution for complex procurement scenarios, demonstrating significant advantages over existing technologies.

[0029] S2. Receive the encrypted uploaded model parameters, decrypt them, use the FedAvg algorithm to obtain the model parameters, update the model parameters through the broadcast update mechanism, use the GraphSAGE model to generate node embedding vectors, calculate the comprehensive risk score using the updated model parameters, and update the knowledge graph by combining the node embedding vectors with the comprehensive risk score. Specifically, the encrypted model parameters are received, decrypted, and then processed using the FedAvg algorithm to obtain the model parameters. The model parameters are then updated via a broadcast update mechanism. This refers to the Central Server updating the encrypted model parameters from K local nodes (local clients, such as procurement systems in different regions or departments, each with its own independent supplier and procurement data). Decryption is performed using a weighted average calculated based on the FedAvg algorithm to determine the global model parameters. The formula is: , , in, This represents the local data volume of the k-th node (approximately 100 procurement events). K represents the total amount of data across all nodes, and K represents the total number of nodes. The central server asynchronously broadcasts the updated global model parameters to all participating local nodes over the network, using message queues, peer-to-peer push, or streaming communication. This allows some nodes to receive updates with acceptable latency, improving flexibility and fault tolerance. Once each local node receives the model parameters... The current local model parameters are replaced and used as the starting point for the next round of training. Local optimization continues, forming a loop of "local training → encrypted upload → central aggregation → broadcast update" that iterates continuously until the model performance converges. This achieves collaborative optimization of the global model while protecting data privacy, and outputs the updated global FM model parameters. .

[0030] The system decrypts encrypted model parameters from K local nodes via a central server and uses the FedAvg algorithm to calculate global model parameters using a weighted average, enabling collaborative optimization of multi-party data. Simultaneously, Paillier homomorphic encryption ensures privacy and security during parameter sharing, overcoming the privacy leakage risks associated with sensitive data sharing in traditional methods. An asynchronous broadcast update mechanism efficiently distributes global model parameters to local nodes via message queues, peer-to-peer push, or streaming communication, supporting flexible updates by some nodes within acceptable latency, thus improving the system's fault tolerance and update efficiency. The iterative "local training → encrypted upload → central aggregation → broadcast update" mechanism ensures continuous convergence of model performance, enhancing the recommendation system's adaptability to distributed procurement scenarios. This provides an efficient and privacy-secure solution for complex supply chain management, demonstrating significant advantages over existing technologies.

[0031] Furthermore, the GraphSAGE model is used to generate node embedding vectors, and the comprehensive risk score is calculated by updating the model parameters. The node embedding vectors are then combined with the comprehensive risk score to update the knowledge graph index based on the initial knowledge graph. Based on this, supplier node features (such as quality rating and historical performance) are fused with contextual information (such as event features and interaction weights) provided by the global FM model and used as input to the GraphSAGE model. Through the model's two-layer aggregation mechanism, neighbor information is aggregated layer by layer and node embeddings are updated. During training, node embeddings are used to generate recommendation ratings, and the BPR loss function of the FM model is used as a supervision signal. By minimizing the ranking error of positive and negative sample pairs, the parameters of GraphSAGE are optimized through backpropagation to obtain a trained GraphSAGE model. Based on the initial knowledge graph By accumulating new events over time (such as new purchases, delivery anomalies, and external risks), an intermediate state graph that has not yet been embedded and updated is generated. Using the trained GraphSAGE model to analyze the intermediate state graph Perform node embedding updates to obtain the latest node embedding vector. The formula is: , in, Let N(s) represent the embedding vector of supplier s at layer l, and N(s) represent the neighboring nodes (upstream and downstream suppliers) in the graph. B is a trainable weight matrix (initially random). It is the ReLU activation function. Let j be the embedding vector of the neighbor node j at layer l. This is the set of neighbors (including upstream and downstream partners) of supplier s in the knowledge graph. Extract interaction weights on key features (such as quality, cost, risk, and ESG) from the global FM model. The weighted average of the feature interaction weights is calculated as the supplier's overall dynamic weight. Dynamic weights The numerical value is obtained by normalization using the Sigmoid function. This represents the current recommendation credibility, and the original edge weights are... (The relationship strength was calculated and normalized using statistical aggregation methods based on historical transaction frequency, contract amount, and cooperation duration business data.) Perform weighted multiplication fusion to obtain the merged edge weights. This enables dynamic adjustment of the importance of edges in the graph; The overall risk score is calculated and propagated using the sentiment score of external events extracted through the VADER sentiment analysis method. The formula is: , in, The risk value of neighbor node j (converted from VADER sentiment score); The latest embedding vector This serves as a new feature vector for suppliers, while simultaneously incorporating the comprehensive risk score calculated from risk propagation. As node attributes, new feature vectors and node attributes are batch-written to the corresponding supplier nodes through the incremental update interface of the Neo4j graph database. For newly added or changed relationships (such as new cooperation events), the edge set (determined by structured data such as contract relationships, delivery transactions, and hierarchical relationships extracted from the SAP ERP system) and the merged edge weights are updated synchronously. After data injection is completed, the graph version management mechanism is triggered, and the current state is updated. Save as a historical snapshot and generate a new knowledge graph. .

[0032] By integrating supplier node features with contextual information from the global FM model based on an initial knowledge graph, and utilizing GraphSAGE's two-layer aggregation mechanism to generate node embedding vectors, this approach effectively captures complex relationships and dynamic features within the supply chain network. It overcomes the shortcomings of traditional methods in semantic modeling and dynamism. Combined with VADER sentiment analysis to extract sentiment scores from external events, it calculates and propagates a comprehensive risk score. Furthermore, by dynamically adjusting edge weights and integrating the interaction weights of the global model, it achieves real-time updates of the knowledge graph and rapid integration of risk information. This significantly improves the recommendation system's responsiveness to dynamic events such as new purchases, delivery anomalies, and external risks. Through Neo4j's incremental update interface and version management mechanism, it ensures the efficiency and consistency of graph updates, solving the problem of lagging updates in traditional knowledge graphs. This enhances the dynamic adaptability and accuracy of the recommendation system and provides an efficient and real-time supplier recommendation solution for complex supply chain scenarios.

[0033] S3. Based on the updated knowledge graph, obtain the updated risk attributes. Combine the risk attributes with the predicted recommendation score to obtain the final recommendation score of the supplier. Generate a supplier recommendation list based on the recommendation score. Calculate the recommendation hit rate using the supplier recommendation list. Update the recommendation strategy based on the recommendation hit rate to achieve closed-loop adaptation. Specifically, the updated risk attributes are obtained based on the updated knowledge graph. These risk attributes are then combined with the predicted recommendation score to obtain the final supplier recommendation score. Finally, a supplier recommendation list is generated based on this recommendation score, which is based on the updated knowledge graph. Extract the latest node information, including the 32-dimensional embedding vector obtained using the GraphSAGE model inference. As a structured representation of suppliers, it captures their roles and contextual relationships within the supply chain network, and obtains updated risk attributes. It reflects the impact of external events; Fusion prediction recommendation score Combined with contextual information enhanced by graph embedding and dynamic risk scores, the final recommendation score for supplier s is calculated. The formula is: , in, To achieve risk avoidance, the risk score is changed from "negative" to "positive". The higher the risk, the lower the value and the lower the recommendation score. Conversely, the lower the risk, the higher the value and the more favorable the recommendation. This achieves the effect of automatically avoiding high-risk suppliers. The fusion weight γ is obtained by Q-learning optimization. For example, γ=0.7 ensures that the recommendation is balanced between accuracy and robustness. All candidate suppliers are ranked by recommendation score The top 10 suppliers are selected in descending order to form a Top-10 recommendation list. The output includes supplier ID and comprehensive score. By extracting embedding vectors from the updated knowledge graph using GraphSAGE model inference as a structured representation of the suppliers, and combining them with dynamic risk attributes obtained from VADER sentiment analysis, the system can accurately capture the roles, contextual relationships, and external event impacts of suppliers in the supply chain network. This overcomes the shortcomings of traditional methods in responding to dynamic risks. By fusing predicted recommendation scores with risk attributes to obtain the final recommendation score, a balance is struck between accuracy and robustness. This enables automatic avoidance of high-risk suppliers and addresses the limitations of traditional single-dimensional recommendations. By generating a Top-10 supplier recommendation list in descending order of recommendation scores, the accuracy, dynamic adaptability, and fairness of the recommendation results are improved, providing an efficient and reliable supplier selection solution for complex procurement scenarios.

[0034] Furthermore, the recommendation hit rate is calculated using the supplier recommendation list. The recommendation strategy is updated based on the hit rate to achieve closed-loop adaptive recommendation. This involves comparing the generated Top-10 recommendation list with the actual selections made by the buyer in the ERP system to calculate the hit rate. Specifically, the generated Top-10 recommended supplier list is compared with the suppliers actually selected by the buyer in the ERP system during the same period. The number of overlaps is counted, and then the overlap is divided by 10 to obtain the hit rate. Combined with the buyer's manual rating of the recommendation results (1–5 points, including ESG compliance feedback), a reward signal is constructed and input into the Q-learning reinforcement learning model to dynamically adjust key parameters in the recommendation process, including the fusion weight γ of the comprehensive score and the interaction weights in the factorization machine. The Q-learning reinforcement learning model uses a weighted sum of hit rate and human rating as reward, and optimizes the strategy through a time difference update mechanism to improve the quality and sustainability of future recommendations. The feedback loop is executed once every fixed number of days (e.g., 7 days), and the updated parameters are fed back to the recommendation model. At the same time, all feedback data is stored in the Neo4j graph database to support efficient querying and long-term iterative optimization.

[0035] By extracting embedding vectors from the updated knowledge graph using GraphSAGE model inference, and using them as structured representations of suppliers, combined with dynamic risk attributes obtained from VADER sentiment analysis, the system accurately captures the roles, contextual relationships, and impact of external events of suppliers in the supply chain network. This overcomes the shortcomings of traditional methods in responding slowly to dynamic risks. The system adopts a fusion of predicted recommendation scores and risk attribute final recommendation scores, and achieves a balance between recommendation accuracy and robustness through Q-learning-optimized fusion weights. It automatically avoids high-risk suppliers, solves the problem of traditional single-dimensional recommendation models lacking comprehensiveness and fairness, and significantly improves the accuracy, dynamic adaptability, and fairness of recommendation results.

[0036] This embodiment also provides an intelligent supplier recommendation system based on artificial intelligence, including: The data acquisition and preprocessing module is used to collect supplier data to construct a sparse participation matrix and an initial knowledge graph, and generate a joint input vector. The fuzzy reasoning and rating prediction module is used to calculate quality and risk membership vectors, and combines the Sugeno model and the PROMETHEE method to generate fuzzy interaction weights and predict recommendation ratings. The federated learning and model update module receives encrypted model parameters, aggregates global parameters using the FedAvg algorithm, and achieves collaborative optimization through a broadcast update mechanism. The graph embedding and risk assessment module is used to calculate a comprehensive risk score by combining global model parameters and VADER sentiment analysis, and to update the node features and edge weights of the knowledge graph. The recommendation generation and adaptive optimization module is used to update the knowledge graph, extract risk attributes, integrate prediction scores to generate a supplier recommendation list, calculate the recommendation hit rate, and optimize the recommendation strategy through Q-learning.

[0037] This embodiment also provides a computer device applicable to an artificial intelligence-based intelligent supplier recommendation method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the artificial intelligence-based intelligent supplier recommendation method proposed in the above embodiment.

[0038] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0039] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements an artificial intelligence-based intelligent supplier recommendation method and system as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0040] In summary, this invention constructs a sparse participation matrix and knowledge graph, combining fuzzy inference methods with VADER sentiment analysis to accurately capture supplier quality and dynamic risk characteristics, effectively improving recommendation accuracy and robustness. It utilizes the GraphSAGE model and FedAvg algorithm to achieve dynamic updates of the knowledge graph and privacy protection under multi-party collaboration, addressing the slow response to external events and data privacy leaks inherent in traditional methods. Furthermore, it optimizes recommendation ranking through the BPR loss function and generates a comprehensive preference index using the PROMETHEE and CRITIC methods, ensuring the balance and fairness of recommendation results across multiple dimensions. This significantly improves the accuracy, adaptability, and security of supplier recommendations, fully meeting the needs of complex procurement scenarios.

[0041] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent supplier recommendation based on artificial intelligence, characterized in that: include, By acquiring procurement events and historical supplier participation data, a sparse participation matrix and knowledge graph are constructed, and a joint input vector is generated. Based on the joint input vector, the quality membership vector is calculated, the VADER method is applied to obtain the risk membership vector, and the Sugeno model is combined to calculate the fuzzy interaction weights. The interaction weights are used to generate the predicted recommendation score, and the parameters of the encrypted model are obtained by combining the BPR loss function. The system receives encrypted uploaded model parameters, decrypts them, uses the FedAvg algorithm to obtain the model parameters, updates the model parameters through a broadcast update mechanism, uses the GraphSAGE model to generate node embedding vectors, calculates the comprehensive risk score using the updated model parameters, and updates the knowledge graph by combining the node embedding vectors with the comprehensive risk score. The updated risk attributes are obtained by updating the knowledge graph. The risk attributes are combined with the predicted recommendation score to obtain the final recommendation score of the supplier. A supplier recommendation list is generated based on the recommendation score. The recommendation hit rate is calculated using the supplier recommendation list. The recommendation strategy is updated based on the recommendation hit rate to achieve closed-loop adaptation.

2. The intelligent supplier recommendation method based on artificial intelligence as described in claim 1, characterized in that: The process involves acquiring procurement event and supplier historical participation data to construct a sparse participation matrix and knowledge graph, and generating a joint input vector. This is achieved by collecting procurement event logs, supplier historical participation records, and external events to construct the sparse participation matrix. , which contains matrix elements Constructing an initial knowledge graph When matrix elements At that time, the purchaser ID is one-hot encoded, and the requirement text is converted into a semantic feature vector using TF-IDF. This is then concatenated with the supplier's identifier vector to generate a joint input vector. .

3. The intelligent supplier recommendation method based on artificial intelligence as described in claim 2, characterized in that: The process involves calculating a quality membership vector based on a joint input vector, applying the VADER method to obtain a risk membership vector, and combining this with a Sugeno model to calculate fuzzy interaction weights. These interaction weights are then used to generate a predicted recommendation score. Finally, the BPR loss function is used to obtain the encrypted model parameters. Supplier quality ratings are extracted from the buyer's database, and the raw supplier quality ratings are normalized to obtain the final quality rating. Three triangular fuzzy sets are defined using fuzzy inference methods for each quality rating. Membership degree is obtained using trigonometric membership functions. Calculate the quality membership vector Text information is generated using collected external event data related to suppliers. For each text Calculate sentiment scores and generate risk values. Quality rating for each supplier and risk value The difference between the actual value and the predicted value is calculated to obtain the scalar residual. and scalar residual Substitute the values ​​into the three triangular fuzzy sets respectively, and output the membership values ​​in each fuzzy set, relative to the scalar residuals. Applying the Gaussian fuzzy membership function, the fuzzy membership vector is obtained. ; Calculate the interaction weights using the Sugeno fuzzy inference model. The PROMETHEE model is used to generate pairwise comparisons of preference strengths, resulting in a comprehensive preference index. Based on interaction weights and comprehensive preference index Combined with the FM framework, predict recommendation scores Based on predictive recommendation rating Calculate the loss value of the BPR loss function. Calculate the loss function For the gradient of each parameter, update the model parameters, and encrypt only the subset of parameters that need to be shared using the Paillier homomorphic encryption algorithm.

4. The intelligent supplier recommendation method based on artificial intelligence as described in claim 3, characterized in that: The received encrypted model parameters are decrypted and then processed using the FedAvg algorithm to obtain the model parameters. The model parameters are then updated via a broadcast update mechanism, indicating that the central server is updating the encrypted model parameters from K local nodes. Decryption is performed using a weighted average calculated based on the FedAvg algorithm to determine the global model parameters. The central server asynchronously broadcasts the updated global model parameters to all participating local nodes over the network, using message queues, peer-to-peer push, or streaming communication. This allows some nodes to receive the updates within an acceptable delay. Once each local node receives the model parameters... Replace the current local model parameters as the starting point for the next round of training, continue local optimization until the model performance converges, and output the updated global FM model parameters. .

5. The intelligent supplier recommendation method based on artificial intelligence as described in claim 4, characterized in that: The process of using the GraphSAGE model to generate node embedding vectors, calculating a comprehensive risk score by updating model parameters, and then updating the knowledge graph using the node embedding vectors and the comprehensive risk score refers to updating the knowledge graph based on the initial knowledge graph. Generate intermediate state graphs that have not yet been embedded and updated. Using the trained GraphSAGE model to analyze the intermediate state graph Perform node embedding updates to obtain the latest node embedding vector. Calculate the edge weights after fusion The overall risk score is calculated and propagated based on the sentiment score of external events extracted using the VADER sentiment analysis method. , the latest embedding vector This serves as a new feature vector for suppliers, while simultaneously incorporating the comprehensive risk score calculated from risk propagation. As node attributes, the new feature vectors and node attributes are batch-written into the corresponding supplier nodes to generate a new knowledge graph. .

6. The intelligent supplier recommendation method based on artificial intelligence as described in claim 5, characterized in that: The process of obtaining updated risk attributes based on an updated knowledge graph, combining these risk attributes with a predicted recommendation score to obtain a final supplier recommendation score, and generating a supplier recommendation list based on the recommendation score refers to the process of using an updated knowledge graph. Extracting the latest node information, including obtaining the embedding vector using the GraphSAGE model inference. As a structured representation of the supplier, obtain the updated risk attributes. Fusion prediction recommendation score Calculate the supplier recommendation score by combining contextual information enhanced by graph embedding and dynamic risk scores. The top 10 items were selected to form the Top-10 recommendation list.

7. The intelligent supplier recommendation method based on artificial intelligence as described in claim 6, characterized in that: The process involves calculating the recommendation hit rate using a supplier recommendation list and updating the recommendation strategy based on the hit rate to achieve closed-loop adaptive learning. This involves comparing the generated Top-10 recommendation list with the actual selections made by the purchaser in the ERP system, calculating the recommendation hit rate, and combining the purchaser's human rating of the recommendation results to construct a reward signal. This signal is then input into a Q-learning reinforcement learning model. The Q-learning reinforcement learning model uses the weighted sum of the hit rate and human rating as the reward and optimizes the strategy through a time-difference update mechanism.

8. An intelligent supplier recommendation system based on artificial intelligence, based on the intelligent supplier recommendation method based on artificial intelligence as described in any one of claims 1 to 7, characterized in that: include, The data acquisition and preprocessing module is used to collect supplier data to construct a sparse participation matrix and an initial knowledge graph, and generate a joint input vector. The fuzzy reasoning and rating prediction module is used to calculate quality and risk membership vectors, and combines the Sugeno model and the PROMETHEE method to generate fuzzy interaction weights and predict recommendation ratings. The federated learning and model update module receives encrypted model parameters, aggregates global parameters using the FedAvg algorithm, and achieves collaborative optimization through a broadcast update mechanism. The graph embedding and risk assessment module is used to calculate a comprehensive risk score by combining global model parameters and VADER sentiment analysis, and to update the node features and edge weights of the knowledge graph. The recommendation generation and adaptive optimization module is used to update the knowledge graph, extract risk attributes, integrate prediction scores to generate a supplier recommendation list, calculate the recommendation hit rate, and optimize the recommendation strategy through Q-learning.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent supplier recommendation method based on artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent supplier recommendation method based on artificial intelligence as described in any one of claims 1 to 7.

Images