Method and apparatus for multi-path recall and ranking fusion of vendor recommendations

Abstract

The application provides a multi-path recall and ranking fusion supplier recommendation method and device, and relates to the technical field of intelligent recommendation. The method aims to solve the problem that the existing technology cannot accurately understand complex material information and cannot balance recommendation coverage and accuracy. The method comprises the following steps: determining the target leaf class of the material by using a text entity recognition model and a material multi-label classification model; performing multi-path supplier recall to obtain a candidate supplier set and calculate a recall score; inputting a constructed ranking feature vector into a ranking model to obtain a transaction probability; obtaining a supplier portrait score of the candidate supplier in the target leaf class; and weighting and fusing the transaction probability, the recall score and the portrait score through a scoring function, calculating a comprehensive score and outputting the score in sequence. The application effectively improves the accuracy and coverage of the recommendation and enhances the interpretability of the scheme by deeply analyzing the material, multi-path recall and three-dimensional information fusion.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology, specifically to a supplier recommendation method that integrates multi-path recall and ranking. Background Technology

[0002] In specialized business scenarios such as industrial product procurement, buyers need to quickly and accurately select suitable partners from a massive supplier database based on specific material requirements. Existing supplier recommendation technologies typically employ a two-stage architecture of "recall-ranking." In the recall stage, some methods use a multi-path recall strategy, such as comprehensively utilizing multiple algorithms based on historical transaction data and content information to recall a subset of candidates, and then weighting and combining the results of each recall to improve the breadth of recommendations.

[0003] However, existing technologies still have many shortcomings. First, descriptions of industrial materials are typically unstructured long texts, mixing various languages, numbers, and symbols, and containing multi-dimensional key information such as model, specifications, material, brand, and technical parameters. Traditional keyword matching or simple word segmentation technologies struggle to accurately and completely capture this complex semantic information, leading to biases in the understanding of material requirements and consequently affecting the accuracy of subsequent recommendations. Second, different companies often use custom, non-standard material classification systems, making recommendation methods relying on universal classification systems difficult to apply directly and limiting their applicability. Third, historical transaction data from suppliers generally suffers from data sparsity and a "long tail effect," meaning a few popular materials have numerous transaction records, while many long-tail materials have very few. Relying solely on historical transaction data for recommendations makes it difficult to recall a sufficient number of candidate suppliers for new or less popular materials, resulting in insufficient recommendation coverage. Finally, in existing technologies, the recall layer and the ranking layer are often disconnected. Simply merging or weighting the results of multiple recalls fails to fully utilize the ranking model's accurate predictive ability of the likelihood of a transaction, making it difficult to effectively balance the accuracy and coverage of recommendations. Furthermore, some solutions based on complex models often lack interpretability in their recommendation results, making it difficult to meet the requirements of transparency and credibility in corporate procurement decisions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a supplier recommendation method and apparatus that integrates multi-path recall and ranking.

[0005] A supplier recommendation method integrating multi-path recall and ranking, provided by the present invention, includes: When receiving a request for a quote to be recommended The material description text in the inquiry is identified by using a pre-trained text entity recognition model, and the target leaf class of the material is determined by combining the material-leaf class relationship data and the pre-trained material multi-label classification model. Execute multi-path supplier recall, including recall paths based on historical transactions and recall paths based on material similarity, to obtain a set of candidate suppliers. Then, for each candidate supplier obtained through each recall path, perform a weighted score according to the preset weight of each path to obtain a recall score m for each candidate supplier. Construct a sorting feature vector containing characteristics of the inquirer, characteristics of the supplier, and characteristics of the material leaf class, and input the sorting feature vector into a pre-trained sorting model to obtain the supplier's transaction probability p under the inquiry; Obtain the supplier profile score l of the candidate supplier on the target leaf class, perform weighted fusion of the transaction probability p, the recall score m and the supplier profile score l to calculate the comprehensive score of the candidate supplier, sort the candidate suppliers according to the comprehensive score, and output the recommendation result.

[0006] Preferably, the multi-path supplier recall specifically includes: The methods include direct historical supplier recall, leaf-type supplier recall, keyword-based material similarity recall, multi-field vector material similarity recall, and supplier similarity extended recall.

[0007] Preferably, the multi-field vector material similarity recall includes: Vectorize the material name, model specification, and technical attribute text respectively, and build independent indexes; Then, the similarity of each field is weighted and fused.

[0008] Preferably, the supplier profile score l is a supplier concentration score calculated based on leaf TF-IDF; Furthermore, the weighted fusion is achieved through the following preset scoring function: , in, The preset fusion weight coefficient is greater than zero.

[0009] Preferably, the pre-trained ranking model is selected from a group consisting of a generalized linear model and a Wide&Deep model.

[0010] Preferably, the multi-label material classification model incorporates a focus loss into its loss function during training.

[0011] Preferably, the target leaf type for determining the material includes: Prioritize searching in the pre-built material-leaf relationship table; If no match is found, the material multi-label classification model is used for prediction, and the prediction results are backfilled into the material-leaf relationship table.

[0012] Preferably, the step of performing multi-path supplier recall adopts a tiered dynamic triggering strategy: First, execute the preset first-level recall path to determine whether the number of candidate suppliers recalled has reached the preset threshold. If the target is not met, the subsequent recall paths will be triggered.

[0013] Preferably, the values ​​of the fusion weighting coefficients α, β, and γ are dynamically determined based on the contextual analysis results of the current inquiry; The context analysis includes at least determining whether the current inquiry belongs to a cold start scenario.

[0014] A supplier recommendation device according to the present invention includes: The inquiry parsing module is used to perform entity recognition on the material description text in the inquiry using a pre-trained text entity recognition model, and combine the material-leaf class relationship data and the pre-trained material multi-label classification model to determine the target leaf class of the material. The recall module is used to execute multi-path supplier recall, including recall paths based on historical transactions and recall paths based on material similarity, to obtain a set of candidate suppliers. The candidate suppliers obtained through each recall path are weighted and scored according to the preset weights of each path to obtain a recall score m for each candidate supplier. The prediction module is used to construct a sorting feature vector containing the characteristics of the inquirer, the characteristics of the supplier, and the characteristics of the material leaf class, and input the sorting feature vector into a pre-trained sorting model to obtain the supplier's probability of closing the deal under the inquiry. The scoring and ranking module is used to obtain the supplier profile score l of the candidate supplier on the target leaf class, perform weighted fusion of the transaction probability p, the recall score m and the supplier profile score l to calculate the comprehensive score of the candidate supplier, and rank the candidate supplier according to the comprehensive score to output the recommendation result.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. Improved the accuracy of material understanding. By adopting a text entity recognition model and a multi-label material classification model, it can accurately extract key information from complex and unstructured industrial material descriptions and adapt it to the enterprise's customized classification system, laying a solid foundation for subsequent accurate recommendations.

[0016] 2. Improved the coverage of recommendations. By integrating multiple recall mechanisms such as historical direct supply, leaf supply, material similarity and supplier similarity, the problem caused by data sparsity and cold start was effectively alleviated, and the breadth and diversity of the candidate supplier set were significantly improved.

[0017] 3. This application creatively proposes a strategy of weighted fusion of recall score in the recall stage, prediction probability in the ranking stage, and supplier static profile score. This strategy effectively balances the "breadth" (reflected by recall score), "accuracy" (reflected by transaction probability), and "business relevance" (reflected by supplier profile score) of the recommendation system, making the final recommendation results more accurate and reasonable, and taking into account both the accuracy and coverage of the recommendation.

[0018] 4. The interpretability and flexibility of the solution are enhanced. By setting configurable weight coefficients for the recall path and the final fusion score, business personnel can flexibly adjust the strategy emphasis according to specific scenarios, so that the recommendation process is no longer a "black box" and the interpretability and controllability of the results are enhanced. Attached Figure Description

[0019] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A functional block diagram of the recommendation system provided in the embodiments of this application; Figure 2 A flowchart illustrating the supplier recommendation method provided in this application embodiment.

[0020] Explanation of reference numerals in the attached figures: 10. Data Model Layer; 11. Entity Recognition Model; 12. Material Classification Model; 13. Supplier Ranking Model; 14. Material and Supplier Inverted Index; 20. Online Service Layer; 21. Inquiry Semantic Analysis Module; 22. Multi-Path Recall Module; 23. Supplier Fine-Grained Ranking Module; 30. Client; 40. Recommendation Server; 41. Text Parsing Module; 42. Recall Module; 43. Ranking Module; 44. Fusion Module. Detailed Implementation

[0021] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0022] Example 1 This application provides a supplier recommendation method that integrates multi-path recall and ranking. In professional scenarios such as industrial product procurement, the coverage and accuracy of supplier recommendations are often insufficient due to issues such as complex material descriptions, non-standard classification systems, and sparse historical data. This application aims to solve these technical problems. In the technical solution of this embodiment, by deeply analyzing material information, combining multi-path recall, precise ranking, and supplier profiling, and finally through an innovative fusion strategy, a comprehensive improvement in recommendation performance is achieved.

[0023] Please see Figure 1 This diagram illustrates a functional block diagram of a recommendation system according to an embodiment of this application. Logically, the system can be divided into a data model layer 10 and an online service layer 20. The data model layer 10 stores various offline models and preprocessed data required for system operation, serving as the foundation for the online service. Correspondingly, the online service layer 20 is responsible for responding to user inquiry requests in real time and executing the complete recommendation process.

[0024] Specifically, the data model layer 10 may include an entity recognition model 11, a material classification model 12, a supplier ranking model 13, and a series of data indexes to support rapid retrieval. The entity recognition model 11 is used to extract key information from unstructured material description text; the material classification model 12 is used to map materials to a company-defined classification system; and the supplier ranking model 13 is used to predict the probability of a specific supplier fulfilling a specific inquiry. An example of the data index is a material-supplier inverted index 14, which records suppliers who have historically supplied a specific material. Furthermore, the data model layer 10 may also include a material vector model, a supplier vector model, a leaf-type supplier inverted index, and supplier leaf-type term frequency-inverse document frequency profile data, etc.

[0025] The online service layer 20 includes an inquiry semantic analysis module 21, a multi-path recall module 22, and a supplier fine-grained ranking module 23. The inquiry semantic analysis module 21 receives and initially parses user inquiries; the multi-path recall module 22, based on the semantic analysis results, uses various strategies in parallel or sequentially to filter out a preliminary candidate set from a massive supplier database; and the supplier fine-grained ranking module 23 performs final scoring and ranking of the candidate suppliers. It should be noted that the supplier fine-grained ranking module 23 is one of the core innovations of this application; it not only performs ranking but also integrates information from the recall phase and supplier profiles.

[0026] The specific process of the supplier recommendation method provided in this embodiment can be found in [reference]. Figure 2 As shown. Figure 2 The flowchart illustrates the supplier recommendation method provided in this application embodiment, showing in detail the various steps of the online recommendation stage.

[0027] This method includes an offline modeling phase and an online recommendation phase.

[0028] During the offline modeling phase, the system first needs to perform thorough data preparation and model training. The data sources mainly include the company's internal historical inquiry data (including material description text), historical order data (including completed transactions with suppliers and materials), a material basic information database, and a supplier information database.

[0029] The specific process of offline modeling includes: 1. Training a text entity recognition model 11. To accurately understand the complex descriptions of industrial materials, it is necessary to construct an entity labeling system for this field. For example, entity labels such as "material body," "model," "material," "standard," "technical parameters," and "brand" can be defined. Then, a batch of representative material description texts are manually labeled to form a training dataset. Based on this labeled dataset, a pre-trained language model, such as the BERT model based on the Transformer architecture, can be selected and fine-tuned to obtain a text entity recognition model 11 that can accurately recognize the above entities. 2. Training a multi-label material classification model 12. Since the material classification systems (usually called leaf classes) of different enterprises are different and there is a long-tail effect, this embodiment adopts the idea of ​​multi-label classification. Historical material samples are associated with their corresponding enterprise-defined leaf class labels to form a multi-label training set. In order to solve the problem of data imbalance, that is, the problem of too many samples of a few popular leaf classes and too few samples of a large number of long-tail leaf classes, a focus loss can be introduced into the loss function when training the model. It should be noted that the focusing loss, through a modulating factor, reduces the contribution of easily classified samples (usually high-frequency leaf class samples) to the overall loss, allowing the model to focus more on difficult-to-classify samples (usually long-tailed leaf class samples). Its loss function can be expressed as: ,in It is the model's prediction of the probability that a material belongs to a certain leaf class. It is the true label (0 or 1) of the leaf class. It is a moderating factor greater than 0. The material classification model 12 trained in this way can more accurately predict the leaf class to which a new or unpopular material belongs. 3. Construct various indexes and profiles. Based on historical transaction data, construct a material-supplier inverted index 14 to quickly find historical suppliers of a certain material. At the same time, construct a leaf class-supplier inverted index to find all suppliers under a certain leaf class. In addition, it is necessary to construct a profile for each supplier under different leaf classes. In this embodiment, the term frequency-inverse document frequency algorithm is used to calculate the supplier profile score. Specifically, for a supplier, the more times it supplies under a certain leaf class, the stronger its correlation with that leaf class (corresponding to term frequency TF); at the same time, if a leaf class is supplied by more suppliers, it indicates that the leaf class has lower discrimination (corresponding to inverse document frequency IDF). Therefore, the term frequency-inverse document frequency score of a supplier in a certain leaf class can measure the supplier's focus or expertise in that leaf class. 4. Train the supplier ranking model 13. A training dataset is constructed using historical transaction records as positive samples (labeled 1) and unsuccessful inquiry-supplier pairs as negative samples (labeled 0). The feature vector can include inquirer features (e.g., industry, historical purchasing preferences), supplier features (e.g., region, main business, historical transaction rate), and material category features. In this embodiment, the supplier ranking model 13 can adopt a generalized linear model, which has advantages such as simplicity, fast training speed, and strong interpretability of results, and can learn the linear relationship between various features and the probability of a transaction.

[0030] Once offline modeling is complete, the system can proceed to the online recommendation phase to process real-time supplier recommendation requests. (See reference...) Figure 2 As shown.

[0031] The online recommendation process is as follows: Step S101: Receive inquiry request. Client 30, such as the front-end interface of a procurement management system, sends an inquiry request to recommendation server 40. This request usually contains one or more materials to be purchased, each material is accompanied by a descriptive text, such as: "Purchase a batch of SUS304 stainless steel bolts, M8*20, conforming to GB / T5783 standard".

[0032] Step S102: Analyze the material using the entity recognition model. After receiving the request, the recommendation server 40 calls its internal text parsing module 41 (corresponding to...). Figure 1The semantic analysis module 21 of the inquiry form (in the text) uses a pre-trained text entity recognition model 11 to process the material description text "SUS304 stainless steel bolts, M820, conforming to GB / T5783 standard" and outputs structured entity recognition results, such as: {Material body: "bolt", material: "SUS304", model: "M820", standard: "GB / T5783"}. It is understandable that this structured information, compared to the original text, greatly improves the depth and accuracy of the computer's understanding of material requirements.

[0033] Step S103: Determine the target leaf class. After completing entity recognition, the text parsing module 41 also needs to determine the target leaf class to which the material belongs. In one embodiment of this application, a strategy of "table lookup first, model prediction supplement" is adopted. First, the system attempts to search for the material in a pre-built material-leaf class relationship table (indexed by the material's unique identifier or the hash value of the material description text). If a historical record is found, for example, if "SUS304 stainless steel bolt M8*20" has previously been determined to belong to the "fastener" leaf class, then this leaf class is directly used as the target leaf class. If no record is found (for example, this is a new material), then the pre-trained material multi-label classification model 12 is called for prediction. The model outputs one or more possible leaf classes and their confidence scores based on the material description text, and the system selects the leaf class with the highest confidence score (for example, "fastener") as the target leaf class. In order to achieve knowledge accumulation and learning, the system will also fill the material-leaf class relationship table with the prediction results so that it can quickly query when encountering the same or similar materials next time, thereby improving efficiency. The text parsing module 41 passes the parsed material entity and target leaf information to subsequent modules.

[0034] Step S104: Perform multi-path recall. The recall module 42 of recommended server 40 (corresponding to...) Figure 1The multi-path recall module 22) receives material information and, in order to maximize the discovery of potential qualified suppliers and alleviate data sparsity and cold start problems, executes multiple recall strategies in parallel or serially to form a broad set of candidate suppliers. As an optional implementation, the recall strategy may include the following five paths: 1. Direct historical supplier recall: Using the unique identifier of the material, directly query the material-supplier inverted index 14 to recall all suppliers who have historically supplied the material. This is the most direct and accurate recall method. 2. Leaf category supplier recall: Using the target leaf category (e.g., "fasteners") determined in step S103, query the leaf category-supplier inverted index to recall all suppliers who have supplied materials under the "fasteners" leaf category. This method can expand the recall scope, especially suitable for recommending new materials. 3. Keyword-based material similarity recall: Using the key entities extracted in step S102 (e.g., "bolts", "SUS304", "M8*20") as keywords, perform text retrieval in the material information database to find other materials with similar descriptions, and then recall the historical suppliers of these similar materials as candidate suppliers. 4. Multi-Field Vector Material Similarity Recall: To more accurately measure the similarity between materials, this approach processes material descriptions in greater detail. Specifically, material information is broken down into multiple fields such as "material name," "model specifications," and "technical attributes," and pre-trained word vector models (such as Word2Vec or BERT) are used to vectorize the text of each field, building an independent vector index library for each field. During recall, the corresponding field vectors of the materials to be recommended are searched for similarity in their respective index libraries to obtain a list of similar materials and similarity scores for each field. Then, the similarity of each field is weighted and fused according to preset field weights to calculate the comprehensive similarity between materials, and finally, the historical supplier of the material with the highest comprehensive similarity is recalled. 5. Supplier Similarity Extended Recall: Building upon the suppliers recalled through the above four approaches, a pre-trained supplier vector model is used to find the nearest neighbor suppliers in the vector space for each recalled supplier. These nearest neighbor suppliers are usually similar to the recalled suppliers in terms of business scope and supply capabilities; adding them to the candidate set further expands the diversity of the recall.

[0035] Step S105: Calculate the recall score m. The recall module 42 merges and deduplicates suppliers recalled through all paths to obtain the final set of candidate suppliers. To retain information from the recall phase in subsequent fusion and ranking, a recall score m needs to be calculated for each candidate supplier. This score is calculated based on which paths the supplier was recalled through and the preset weights of each path. For example, weights can be set for five different recall paths. If a supplier A is hit by both direct historical recall and leaf-type recall, its recall score is... It can be This weighted scoring method allows the recall score m to reflect the relevance and source diversity of a candidate supplier to the current inquiry. The recall module 42 ultimately outputs a list of candidate suppliers and the recall score m for each supplier.

[0036] Step S106: Construct the ranking feature vector. The ranking module 43 of the recommendation server 40 (corresponding to...) Figure 1 (Part of the refined supplier ranking module 23) After receiving the list of candidate suppliers, it needs to construct a ranking feature vector for each "inquirer-candidate supplier-material" combination to accurately predict the probability of a transaction. This vector is a high-dimensional vector containing multiple aspects of information that can affect the probability of a transaction, such as: inquirer characteristics (e.g., industry, historical purchase amount, purchase frequency), supplier characteristics (e.g., registered capital, location, number of historical collaborations, positive feedback rate), material leaf category characteristics (e.g., the popularity of the current material's leaf category), and the cross-features between the inquirer and the supplier (e.g., the matching degree between "inquirer industry" and "supplier's main business").

[0037] Step S107: Input the ranking model to predict the transaction probability p. The ranking module 43 inputs the constructed ranking feature vector into the pre-trained supplier ranking model 13 (a generalized linear model in this embodiment). Based on the learned feature weights, the model outputs a value between 0 and 1, which is the predicted transaction probability p. In other words, this probability p represents the likelihood that the supplier will successfully win the bid in this inquiry, and is the core indicator for measuring the "accuracy" of the recommendation.

[0038] Step S108: Obtain the supplier's TF-IDF score. Recommend the fusion module 44 of server 40 (corresponding to...). Figure 1 Another part of the supplier fine-grained ranking module 23 introduces static profile information of suppliers. Based on the target leaf class determined in step S103, it queries the pre-calculated supplier leaf class term frequency-inverse document frequency profile data to obtain the supplier profile score l for each candidate supplier in the target leaf class. As mentioned above, this score l represents the supplier's "professionalism" or "concentration" in the material field.

[0039] Step S109: Calculate the comprehensive score. As a key step in this application, the fusion module 44 weights and fuses three important pieces of information from different dimensions to calculate the final comprehensive score. These three pieces of information are: the probability of conversion p, representing the "accuracy" of the recommendation; the recall score m, representing the "breadth" and "diversity" of the recommendation; and the profile score l, representing the "professionalism" of the supplier. The fusion is achieved through a preset scoring function: Here, α, β, and γ are preset fusion weight coefficients that are greater than zero, and their relative magnitudes determine the focus of the recommendation strategy. For example, if you want more accurate recommendation results, you can increase the value of α; if you want to cover more potential suppliers, you can increase the value of β.

[0040] Step S110: Sort and output the results. After calculating the comprehensive score for all candidate suppliers, the fusion module 44 sorts them in descending order. Finally, the recommendation server 40 returns the sorted supplier list (e.g., the top 10) as the final recommendation result to the client 30.

[0041] Through the above process, the method in this embodiment can provide a recommendation list that balances accuracy and breadth, and incorporates supplier expertise, for complex industrial material price inquiries, thereby significantly improving the efficiency and quality of procurement decisions.

[0042] Example 2 In another embodiment of this application, the technical solution of Embodiment 1 is optimized. This embodiment aims to demonstrate the replaceability and scalability of the ranking model in this application to adapt to more complex business scenarios. The main difference between this embodiment and Embodiment 1 lies in the selection and training method of the supplier ranking model 13, as well as the transaction probability prediction step in the online recommendation stage. In Embodiment 1, the ranking model adopts a generalized linear model, while in this embodiment, a more powerful Wide&Deep model is adopted.

[0043] Please continue to refer to this. Figure 1 and Figure 2 The overall system architecture and recommendation process in this embodiment are basically the same as in Embodiment 1. The steps of data preparation in the offline modeling stage, training of entity recognition model 11, training of material classification model 12, and construction of various indexes and profiles are all the same as in Embodiment 1. The core difference lies in the training of supplier ranking model 13.

[0044] During the offline modeling phase, when training the supplier ranking model 13, this embodiment employs the Wide & Deep model. This model combines the "memory" capability (i.e., its Wide part) of a generalized linear model with the "generalization" capability (i.e., its Deep part) of a deep neural network. The Wide part excels at learning and remembering simple, direct, and highly interpretable cross-features; the Deep part excels at learning latent, highly abstract deep patterns in the data through multi-layered nonlinear transformations.

[0045] The specific training process is as follows: 1. Feature partitioning: When constructing the ranking feature vector, the features need to be divided into two parts, which are input into the Wide and Deep sides of the model respectively. Typically, discrete features that need to be combined to reflect business logic, such as the combination of "the industry of the inquirer" and "the supplier's main business," are input into the Wide side. Continuous features or categorical features that can be represented as dense vectors, such as material vectors or statistical features of the supplier's historical behavior, are input into the Deep side. 2. Model structure: The Wide part is a generalized linear model. The Deep part is usually a multi-layer feedforward neural network, with each layer containing several neurons, and uses activation functions such as ReLU for non-linear transformation. 3. Joint training: The output of the Wide & Deep model is a weighted sum of the outputs of the Wide and Deep parts, which is then passed through an activation function such as Sigmoid to obtain the final predicted probability. During training, the two parts of the model are jointly trained to optimize a unified loss function, such as the logarithmic loss function.

[0046] In the online recommendation phase, the process is similar to that in Example 1, but in step S107, when predicting the transaction probability p, the system calls the trained Wide&Deep model. Specifically, after constructing the ranking feature vector in step S106, the feature vector is similarly divided into two parts, which are input into the Wide and Deep sides of the deployed Wide&Deep model, respectively. After forward propagation calculation, the model outputs a predicted value, which is the transaction probability p of the "inquirer-candidate supplier-material" combination.

[0047] The subsequent steps S108 (obtaining the supplier's TF-IDF score) and S109 (calculating the overall score) are exactly the same as in Example 1. The final overall score function remains the same:

[0048] By employing the Wide&Deep model, this embodiment can uncover more complex nonlinear and cross-relationships between features than Embodiment 1. For example, it may learn higher-order implicit patterns such as "a buyer in a specific industry" preferring to choose a supplier in a specific region when purchasing "materials of a specific leaf type." Therefore, for scenarios with complex purchasing behavior patterns and significant feature interactions, the predicted transaction probability p in this embodiment will be more accurate, thereby improving the ranking quality and business matching of the final recommendation list. It is understood that the solution of this application allows the supplier ranking model 13 to select from a group of models, including generalized linear models and Wide&Deep models, providing a more flexible and powerful technical solution.

[0049] Example 3 As an optional implementation of Embodiment 1, this embodiment aims to optimize the multi-path recall step ( Figure 2 The execution efficiency of S104 in the example is important, especially when handling massive requests, to balance the recommendation effect with system resource consumption. In the scheme of Example 1, the multi-path recall module 22 can execute all recall paths in parallel. However, in some cases, such as when processing popular materials with rich historical transactions, this approach may cause unnecessary consumption of computing resources.

[0050] To address this issue, this embodiment proposes a hierarchical, dynamically triggered recall strategy. The core idea of ​​this strategy is to divide different recall paths into different priorities based on their precision, recall rate, and computational cost, then execute them in priority order, and dynamically determine whether to trigger subsequent higher-level recalls based on intermediate results.

[0051] The overall process of this embodiment is largely the same as that of Embodiment 1, with the main difference being the specific implementation of step S104.

[0052] The specific tiered dynamic triggering recall strategy is as follows: 1. Tiered Recall Path: The five recall paths mentioned in Example 1 are divided into three levels. This division can be adjusted according to actual business scenarios and experience. A typical division scheme is as follows: * Level 1 (High Priority, Low Cost, High Accuracy): Direct historical supplier recall. This path directly utilizes the inverted index, has the fastest calculation speed, and the recalled suppliers are historically proven, with the strongest relevance. * Level 2 (Medium Priority, Medium Cost): Leaf-type supplier recall, keyword material similarity recall. These two paths expand the recall scope, but the calculation cost and accuracy are between Level 1 and Level 3. * Level 3 (Low Priority, High Cost, High Recall Rate): Multi-field vector material similarity recall, supplier similarity extended recall. These two paths involve complex vector calculations and similarity searches, with the highest calculation cost, but can discover the largest range of potential suppliers, which is especially important for cold start scenarios.

[0053] 1. Dynamic Trigger Logic: Set a threshold N for the number of candidate suppliers, for example, N=50. This threshold represents the size of the candidate set that the system expects to provide for the ranking phase. The recall module 42 executes according to the following logic: First, implement the first-level recall path, namely "direct historical supplier recall".

[0054] Then, it is determined whether the number of candidate suppliers currently being recalled has reached the preset threshold N.

[0055] If the number is greater than or equal to N, it indicates that enough strongly relevant suppliers have been found, and the system can consider the recall task to be basically completed. Therefore, it stops executing the subsequent level of the recall path and directly sends the current candidate set to the next step.

[0056] If the number is less than N, it indicates that relying solely on historical direct suppliers is insufficient to form a sufficient candidate set, and the recall scope needs to be further expanded. In this case, the system will continue to trigger and execute the second-level recall path (leaf-type supplier recall and keyword material similarity recall).

[0057] After completing the second stage, the newly recalled suppliers are merged with the results of the first stage to remove duplicates, and the total number is checked again to see if it reaches the threshold N.

[0058] If the number is still less than N, the third-level recall path (multi-field vector material similarity recall and supplier similarity extended recall) will continue to be triggered and executed until all levels of paths have been executed, or the number of candidate suppliers reaches the threshold N.

[0059] After the recall process is completed, the subsequent steps S105 (calculating the recall score m) and S106 to S110 (sorting and merging) are exactly the same as in Example 1. It should be noted that the calculation of the recall score m is still based on which paths the supplier was hit by and the weight of these paths, and is independent of the execution order of the paths.

[0060] By adopting this hierarchical dynamic triggering strategy, this embodiment significantly optimizes system performance while ensuring a sufficient number of final candidate suppliers. For popular materials, recommendation requests may only require executing the first-level recall, greatly reducing average response time and server load. For less popular or new materials, the system can still execute all recall paths through hierarchical triggering to ensure the coverage and diversity of recommendations. This intelligent resource scheduling method makes the entire recommendation system more efficient and scalable.

[0061] Example 4 This embodiment is another variation of Embodiment 1, aiming to improve the intelligence and adaptability of the final scoring fusion strategy. It is understood that in Embodiment 1, the fusion weight coefficients α, β, and γ used to calculate the comprehensive score are preset fixed values. However, the emphasis on the recommendation results varies in different business scenarios. For example, when recommending suppliers for a completely new material with no historical transaction records (i.e., a cold start scenario), the primary goal is to ensure the breadth and coverage of the recommendations to discover as many potential suppliers as possible; in this case, the recall score m should be more important. Conversely, when recommending suppliers for a regularly procured material with a large amount of historical data, the goal is to "find the right supplier," that is, to prioritize recommending suppliers with good historical performance and a high probability of successful transactions; in this case, the probability of successful transactions p should be more important.

[0062] To enable the recommendation strategy to adapt to different scenarios, this embodiment proposes a context-aware dynamic fusion weight strategy. The core idea is to perform context analysis on the current inquiry request before performing the final score fusion, determine its scenario type, and then use different combinations of fusion weight coefficients (α, β, γ) based on different scenario types.

[0063] The overall process of this embodiment is largely the same as that of Embodiment 1, with the main difference being the implementation method of step S109 (calculating the comprehensive score).

[0064] The specific context-aware dynamic fusion weighting strategy is as follows: 1. Context Scene Recognition: Before executing step S109, the system adds a context analysis step, which is responsible for determining the scenario type of the current inquiry request, including at least determining whether the current inquiry belongs to a cold start scenario. A simple judgment method is to query the material-leaf relationship table and historical transaction records. If no information about the material being inquired about is found in these data, it can be determined as a "cold start scenario"; if the material has rich historical transaction records, it can be determined as a "regular procurement scenario". Of course, more granular scenarios can also be defined, such as "high-value material procurement scenario" and "urgent procurement scenario", and the judgment can be made according to business rules.

[0065] 2. Dynamic Weight Adjustment: The system pre-configures a weight library, which stores combinations of weight coefficients (α, β, γ) for different scenarios. For example: For the "cold start scenario": In this scenario, ranking model 13 may not accurately predict the transaction probability p due to a lack of historical data, while the breadth of recall brought by multi-path recall is more important, and the supplier profile l can also provide valuable reference. Therefore, a weighted combination focusing on the recall score m and the profile score l can be used, for example: α=0.2, β=0.6, γ=0.2. This will allow suppliers recalled by multiple paths (especially those with strong scalability) or those more focused on relevant leaf classes to obtain higher initial scores.

[0066] For the "routine procurement scenario": In this scenario, the system has sufficient historical data, the prediction of ranking model 13 is relatively reliable, and the probability of a successful transaction, p, is the most critical indicator for measuring the supplier matching degree. Therefore, a set of weights that emphasize the probability of a successful transaction, p, can be used, for example: α=0.6, β=0.2, γ=0.2.

[0067] 3. Perform fusion calculation: In step S109, the fusion module 44 first calls the context analysis step to obtain the current inquiry scenario type. Then, it retrieves the (α, β, γ) combination corresponding to the scenario from the weight library. Finally, using this dynamically obtained set of weights, it applies the scoring function. To calculate the overall score for each candidate supplier.

[0068] Through this context-aware dynamic weighting strategy, this embodiment makes the recommendation system more intelligent and flexible. It can automatically adjust its evaluation criteria based on the specific question (whether it's a new or old product), prioritizing recall during cold starts and accuracy during mature procurement. This allows it to provide recommendations that better suit user needs in various scenarios, thus improving the overall user experience and the intelligence level of the recommendation system.

[0069] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A multi-path recall and ranking fusion-based supplier recommendation method, characterized in that, include: When receiving a request for a quote to be recommended The material description text in the inquiry is identified by using a pre-trained text entity recognition model, and the target leaf class of the material is determined by combining the material-leaf class relationship data and the pre-trained material multi-label classification model. Execute multi-path supplier recall, including recall paths based on historical transactions and recall paths based on material similarity, to obtain a set of candidate suppliers. Then, for each candidate supplier obtained through each recall path, perform a weighted score according to the preset weight of each path to obtain a recall score m for each candidate supplier. Construct a sorting feature vector containing characteristics of the inquirer, characteristics of the supplier, and characteristics of the material leaf class, and input the sorting feature vector into a pre-trained sorting model to obtain the supplier's transaction probability p under the inquiry; Obtain the supplier profile score l of the candidate supplier on the target leaf class, perform weighted fusion of the transaction probability p, the recall score m and the supplier profile score l to calculate the comprehensive score of the candidate supplier, sort the candidate suppliers according to the comprehensive score, and output the recommendation result. 2.The multi-path recall and ranking fusion based vendor recommendation method of claim 1, wherein, The multi-path supplier recall specifically includes: Direct historical supplier recall, leaf-type supplier recall, keyword-based material similarity recall, multi-field vector material similarity recall, and supplier similarity extended recall. 3.The multi-path recall and ranking fused vendor recommendation method of claim 2, wherein, The multi-field vector material similarity recall includes: Vectorize the material name, model specification, and technical attribute text respectively, and build independent indexes; Then, the similarity of each field is weighted and fused. 4.The multi-path recall and ranking fusion based vendor recommendation method of claim 1, wherein, The supplier profile score l is the supplier concentration score calculated based on leaf class TF-IDF; Furthermore, the weighted fusion is achieved through the following preset scoring function: ， wherein is a preset fusion weight coefficient greater than zero.

5. The multi-path recall and ranking fused vendor recommendation method of claim 1, wherein, The pre-trained ranking model is selected from a group consisting of a generalized linear model and a Wide&Deep model.

6. The multi-path recall and ranking fused vendor recommendation method of claim 1, wherein, The material multi-label classification model incorporates a focus loss into its loss function during training.

7. The multi-path recall and ranking fused vendor recommendation method of claim 1, wherein, The target leaf type for determining the material includes: Prioritize searching in the pre-built material-leaf relationship table; If no match is found, the material multi-label classification model is used for prediction, and the prediction results are backfilled into the material-leaf relationship table.

8. The multi-path recall and ranking fused vendor recommendation method of claim 1, wherein, The step of executing the multi-path supplier recall adopts a tiered, dynamically triggered strategy: First, execute the preset first-level recall path to determine whether the number of recalled candidate suppliers has reached the preset threshold. If the target is not met, the subsequent recall paths will be triggered. 9.The multi-path recall and ranking fusion based vendor recommendation method of claim 4, wherein, The fusion weight coefficient is dynamically determined according to the context analysis result of the current inquiry. The context analysis includes at least determining whether the current inquiry belongs to a cold start scenario.

10. A provider recommendation apparatus characterized by comprising: include: The inquiry parsing module is used to perform entity recognition on the material description text in the inquiry using a pre-trained text entity recognition model, and combine the material-leaf class relationship data and the pre-trained material multi-label classification model to determine the target leaf class of the material. The recall module is used to execute multi-path supplier recall, including recall paths based on historical transactions and recall paths based on material similarity, to obtain a set of candidate suppliers. The candidate suppliers obtained through each recall path are weighted and scored according to the preset weights of each path to obtain a recall score m for each candidate supplier. The prediction module is used to construct a sorting feature vector containing the characteristics of the inquirer, the characteristics of the supplier, and the characteristics of the material leaf class, and input the sorting feature vector into a pre-trained sorting model to obtain the supplier's probability of closing the deal under the inquiry. The scoring and ranking module is used to obtain the supplier profile score l of the candidate supplier on the target leaf class, perform weighted fusion of the transaction probability p, the recall score m and the supplier profile score l to calculate the comprehensive score of the candidate supplier, and rank the candidate supplier according to the comprehensive score to output the recommendation result.

Images