Recommendation model interpretation method and device based on rules, equipment and medium
Through the rule-based recommendation model interpretation method, the matching subgraphs and evidence in the graph data are extracted and updated, and the fidelity and sparseness of the interpretation of recommended results in the prior art are solved, achieving a more efficient and reliable interpretation effect.
Patent Information
- Application Number
- CN202411882772.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-19
- Publication Date
- 2025-03-04
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing recommendation model interpretation methods have problems with fidelity and sparseness in improving the interpretability of recommendation results, resulting in inaccurate or overly complex interpretation results.
The rule-based recommendation model interpretation method is adopted, and the initial matching subgraph is extracted by obtaining interpretation rules, graph data and user item pairs, the target node and matching path are determined, candidate evidence is formed, and the evidence is updated based on the evidence score and initial score threshold, and high priority evidence is dynamically stored.
It improves the effectiveness and reliability of the recommendation model interpretation, enhances users' trust in model results, reduces computational complexity, improves algorithm efficiency, and dynamically updates evidence through the heap structure, allowing users to quickly obtain the most explanatory evidence.
Smart Images
Figure CN119558417A_ABST
Abstract
Description
Technical Field
[0001] The present application is applicable to the field of model interpretation, and in particular relates to a rule-based recommendation model interpretation method, device, equipment and medium. Background Art
[0002] With the rapid development of artificial intelligence technology, Graph Neural Networks (GNN) have shown great potential in recommendation systems. GNN can use multi-hop relationships in graph data to accurately model user preferences and analyze historical user-item interactions to provide users with personalized recommendations. With the popularity of GNN recommendation models, the interpretability of models has become an increasingly important research direction. In recommendation systems, interpretability means being able to clearly explain why a specific item is recommended to a user. This is crucial for users because it can help users understand the source of recommendation results and enhance their trust in the recommendation system. At the same time, for developers, being able to explain GNN recommendation results is also the key to optimizing and debugging models.
[0003] Existing explanation methods can be divided into two categories: self-explanatory GNN models and post-explanation methods. Self-explanatory GNN models build explanation mechanisms inside the model to generate explanations in real time when the model makes predictions. Although such models can generate meta-paths or explanations for recommendations, they often cannot distinguish which features are decisive and under what conditions recommendations can be made. Post-explanation methods generate explanations after the model makes predictions, and provide explanations for recommendation results by analyzing model outputs and extracting subgraphs or features. However, they also face some challenges. Fidelity and sparsity are two key indicators for measuring the effectiveness of explanations. Many existing methods fail to meet the expectations of practitioners in terms of these two indicators, resulting in explanations that may be inaccurate or too complex. Therefore, how to improve the interpretability of recommendation results of recommendation models has become an urgent problem to be solved. Summary of the invention
[0004] In view of this, embodiments of the present application provide a rule-based recommendation model interpretation method, apparatus, device and medium to solve the problem of how to improve the interpretability of the recommendation results of the recommendation model.
[0005] In a first aspect, an embodiment of the present application provides a rule-based recommendation model interpretation method, the recommendation model interpretation method comprising: Obtaining N interpretation rules, graph data, and user-item pairs to be interpreted in the graph data, wherein the user-item pairs are node pairs representing a mapping relationship between users and items obtained by recommending items to users in the graph data based on a recommendation model, and N is an integer greater than zero; Using any interpretation rule, extracting an initial matching subgraph corresponding to the user-item pair from the graph data, and determining an initial rule score upper bound of the interpretation rule according to node scores of matching nodes in the initial matching subgraph; When the upper bound of the initial rule score reaches the initial score threshold, a target node and all matching paths including the target node are determined from the initial matching subgraph, and each matching path is combined with the interpretation rule to form a candidate evidence; For any candidate evidence, determine the evidence score of the candidate evidence according to the rule score representing the importance of the interpretation rule and the path score representing the importance of the corresponding matching path, traverse all candidate evidences to obtain the evidence score of each candidate evidence, and obtain the target evidence according to the evidence scores of all candidate evidences and the initial score threshold; The target evidence is written into a heap structure, the lowest evidence score in the heap structure is used to update the initial score threshold to obtain an updated score threshold, the target node and the child nodes whose parent nodes are the target node are removed from the initial matching subgraph to obtain an updated matching subgraph, the initial rule score upper bound is updated using the updated matching subgraph to obtain an updated rule score upper bound, and the heap structure is used to store target evidence that meets the conditions; The update score threshold is used as the initial score threshold. If the update rule score upper bound reaches the update score threshold, the update matching subgraph is used as the initial matching subgraph, and the determination of a target node from the initial matching subgraph is returned to be executed, and / or, if the update rule score upper bound does not reach the update score threshold, all interpretation rules are traversed to return all target evidences in the heap structure.
[0006] In a second aspect, an embodiment of the present application provides a rule-based recommendation model interpretation device, the recommendation model interpretation device comprising: an acquisition module, configured to acquire N interpretation rules, graph data, and user-item pairs to be interpreted in the graph data, wherein the user-item pairs are node pairs representing a mapping relationship between users and items obtained by recommending items to users in the graph data based on a recommendation model, and N is an integer greater than zero; a rule score determination module, configured to extract an initial matching subgraph corresponding to the user-item pair from the graph data using any interpretation rule, and determine an upper bound of an initial rule score of the interpretation rule according to node scores of matching nodes in the initial matching subgraph; A rule score judgment module, used for determining a target node and all matching paths including the target node from the initial matching subgraph when the upper bound of the initial rule score reaches the initial score threshold, and forming a candidate evidence with each matching path and the interpretation rule; A target evidence judgment module is used to determine the evidence score of any candidate evidence according to the rule score representing the importance of the interpretation rule and the path score representing the importance of the corresponding matching path, traverse all candidate evidences to obtain the evidence score of each candidate evidence, and obtain the target evidence according to the evidence scores of all candidate evidences and the initial score threshold; An updating module is used to write the target evidence into a heap structure, update the initial score threshold using the lowest evidence score in the heap structure to obtain an updated score threshold, remove the target node and the child nodes whose parent nodes are the target node from the initial matching subgraph to obtain an updated matching subgraph, and use the updated matching subgraph to update the initial rule score upper bound to obtain an updated rule score upper bound. The heap structure is used to store target evidence that meets the conditions; A loop module is used to use the update score threshold as the initial score threshold, and if the update rule score upper bound reaches the update score threshold, use the update matching subgraph as the initial matching subgraph, return to execute the determination of a target node from the initial matching subgraph, and / or, if the update rule score upper bound does not reach the update score threshold, traverse all interpretation rules and return all target evidence in the heap structure.
[0007] In a third aspect, an embodiment of the present application provides a computer device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the recommendation model interpretation method as described in the first aspect when executing the computer program.
[0008] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the recommendation model interpretation method as described in the first aspect is implemented.
[0009] Compared with the prior art, the embodiments of the present application have the following beneficial effects: the present application obtains N interpretation rules, graph data, and user-item pairs to be interpreted in the graph data. For any interpretation rule, if the upper bound of the initial rule score reaches the initial score threshold, a target node and a matching path are determined from the initial matching subgraph, the matching path and the interpretation rule form candidate evidence, the evidence score is determined, and the target evidence is obtained according to the evidence score and the initial score threshold. The target evidence is written into a heap structure, the initial score threshold, the initial matching subgraph, and the upper bound of the initial rule score are updated to obtain an updated score threshold, an updated initial matching subgraph, and an updated rule score upper bound, the updated score threshold is set to the initial score threshold, if the upper bound of the updated rule score reaches the updated score threshold, the updated matching subgraph is set to the initial matching subgraph, and the execution is returned to determine a target node from the initial matching subgraph, and / or, if not reached, all interpretation rules are traversed to return all target evidence in the heap structure.
[0010] Among them, the recommendation results of the recommendation model for recommending items to users in the graph data are explained through interpretation rules. On the basis of the subgraph, the topological structure and node features that have a decisive influence on the recommendation results are extracted to form target evidence, optimize the interpretation method, and further help users understand which features play a key role in the model decision, thereby enhancing users' trust in the model results, improving the effectiveness (fidelity and sparsity) and reliability of the interpretation, and based on the rule score upper bound and the initial score threshold, it ensures that high-priority (higher evidence score) target evidence that has a decisive influence on the recommendation results can be efficiently screened out, reducing the computational complexity and improving the efficiency of the algorithm. The heap structure is maintained to dynamically update and store the current target evidence with high priority (higher evidence score), so that users can quickly obtain the most explanatory and influential evidence, improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0012] Figure 1 This is a schematic diagram of an application environment of a rule-based recommendation model interpretation method provided in Example 1 of the present application; Figure 2 It is a flowchart of a rule-based recommendation model interpretation method provided in Example 2 of the present application; Figure 3 This is a schematic diagram of a graph data provided in Example 2 of the present application; Figure 4 This is a schematic diagram of an interpretation rule provided in Example 2 of the present application; Figure 5 This is a schematic diagram of an initial matching subgraph provided in Example 2 of the present application; Figure 6 This is a schematic diagram of an upper bound of a variable score provided in Example 2 of the present application; Figure 7 This is a schematic diagram of an influence score provided in Example 2 of the present application; Figure 8 It is a flowchart of a rule-based recommendation model interpretation method provided in Example 3 of the present application; Fig. 9 It is a flowchart of a rule-based recommendation model interpretation method provided in Example 4 of the present application; Fig.10 It is a flowchart of a rule-based recommendation model interpretation method provided in Example 5 of the present application; Fig.11 It is a structural diagram of a rule-based recommendation model interpretation device provided in Example 6 of the present application; Fig.12 It is a structural diagram of a computer device provided in Example 7 of the present application. DETAILED DESCRIPTION
[0013] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.
[0014] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.
[0015] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0016] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.
[0017] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.
[0018] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.
[0019] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.
[0020] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.
[0021] It should be understood that the size of the serial numbers of the steps in the following embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.
[0022] In order to illustrate the technical solution of the present application, a specific embodiment is provided below for illustration.
[0023] The rule-based recommendation model interpretation method provided in the first embodiment of the present application can be applied to Figure 1 In the application environment, the server communicates with the client, the server provides the recommendation model explanation service, and the client triggers the recommendation model explanation task to the server. The client includes but is not limited to PDAs, desktop computers, laptops, ultra-mobile personal computers (UMPC), netbooks, cloud computing devices, personal digital assistants (PDA), and other devices. The computer device corresponding to the server can be implemented with an independent server or a server cluster consisting of multiple servers.
[0024] See also Figure 2 , is a flowchart of a rule-based recommendation model interpretation method provided in Example 2 of the present application. The above recommendation model interpretation method is applied to Figure 1 The server in the example connects to the client to obtain N interpretation rules, graph data, and user-item pairs to be interpreted in the graph data sent by the client. Figure 2 As shown, the recommendation model explanation method may include the following steps: Step S201, obtaining N interpretation rules, graph data, and user-item pairs to be interpreted in the graph data.
[0025] In this embodiment, the user-item pair is a node pair representing the mapping relationship between users and items obtained by recommending items to users in the graph data based on the recommendation model, and N is an integer greater than zero; For example, Figure 3 As shown, it is a schematic diagram of a graph data provided by the second embodiment of the present application. is a user-item pair in the graph data, , M is the recommendation model, u represents the user, and v represents the item.
[0026] The basic form of Rules for ExPlanations (REPs) is , where Q is a graph pattern used to depict the topological structure between user x and item y related to the recommendation decision. The set of nodes related to the user can be expressed as , nodes related to items are represented as , when referring to any node in the graph pattern, it can be represented by z, that is, , It is a set of predicates that describe the conditions and features related to node features; For example, Figure 4The figure shows a schematic diagram of an interpretation rule provided in the second embodiment of the present application.
[0027] Interpretation Rules The basic form is ,in, for , interpretation rules Combination Mode and prerequisites , explains the Recommended Movies The reasons are: 1) Watched at least two action movies before (they are similar to of the same type); 2) There is a person who likes high-rated movies (Movie rating ≥ 6) and by Director; 3) Is Director's award-winning film (because ).
[0028] Step S202: using any interpretation rule, extracting an initial matching subgraph corresponding to the user-item pair from the graph data, and determining an upper bound of the initial rule score of the interpretation rule according to the node scores of the matching nodes in the initial matching subgraph.
[0029] Step S203, when the upper bound of the initial rule score reaches the initial score threshold, a target node and all matching paths including the target node are determined from the initial matching subgraph, and each matching path is combined with the interpretation rule to form a candidate evidence.
[0030] In this embodiment, the initial matching subgraph may refer to the matching subgraph of the interpretation rule at the user-item pair in the graph data. For any interpretation rule, the interpretation rule may have multiple matching results at a user-item pair in the graph data, and the initial matching subgraph includes all matching results of the interpretation rule at the user-item pair; the matching node may refer to a node in the matching subgraph, and the node score may refer to a score that represents the importance of the node.
[0031] The initial score threshold may refer to a pre-set score threshold; the matching path may refer to a path in the initial matching subgraph that represents any matching result of an interpretation rule; the target node may refer to any node in the initial matching subgraph; the candidate evidence may be composed of an interpretation rule and any matching path (matching result) of the interpretation rule, which is used to locally explain the user-item pairs obtained by the recommendation model decision.
[0032] Given N explanation rules, a simple algorithm for calculating all target evidences corresponding to the explanation rules is: for each explanation rule, find all matching results (matching paths) with the graph pattern at the user-item pair in the graph data, and form evidence with the matching results (matching paths) that satisfy the explanation rule. However, this algorithm has certain limitations in practical applications: 1) When the graph data is dense, it is very expensive to find all matching results (matching paths) at the user-item pair that can match the explanation rules; 2) Multiple explanation rules may be applicable to the user-item pair, and there may be multiple matching results (matching paths) for each explanation rule; Therefore, this application adopts a pruning strategy to filter out those evidences with low priority (lower evidence score) and finally obtain evidences with high priority (higher evidence score). The core of the pruning strategy is the upper bound of the ranking scores of all evidences involved in the interpretation rules, denoted as: ,in, is the upper bound of the initial rule score, Score the evidence.
[0033] Specifically, given an interpretation rule, if the upper bound of the initial rule score (updated rule score upper bound) corresponding to the interpretation rule does not reach the initial score threshold (updated score threshold), it means that all evidence formed by the interpretation rule and the corresponding matching result cannot contribute to the evidence with high priority (higher evidence score), and the processing of the interpretation rule should be stopped as soon as possible, and the unprocessed interpretation rules in the N interpretation rules should continue to be traversed, and all target evidence with high priority (higher evidence score) in the heap structure should be returned; if the upper bound of the initial rule score (updated rule score upper bound) corresponding to the interpretation rule reaches the initial score threshold (updated score threshold), it means that there is evidence in all evidence formed by the interpretation rule and the corresponding matching result that can contribute to the evidence with high priority (higher evidence score), and the interpretation rule is processed, and the steps of determining a target node and all matching paths containing the target node from the initial matching subgraph are executed, and each matching path is formed into a candidate evidence with the interpretation rule.
[0034] Specifically, when the upper bound of the initial rule score reaches the initial score threshold and the target node is determined, for any variable in the interpretation rule, the impact score of removing any matching node mapped to the variable on the variable score upper bound is calculated, and the matching node with the largest impact score is determined as the target node; after the target node is determined, all matching paths (matching results) containing the target node are determined from the initial matching subgraph, and each matching path is combined with the interpretation rule to form a candidate evidence.
[0035] For example, Figure 5 As shown, it is a schematic diagram of an initial matching subgraph provided in the second embodiment of the present application; Figure 6 , which is a schematic diagram of an upper bound of a variable score provided in Example 2 of the present application; In determining the upper bound of the variable score, Figure 4 Rules of interpretation shown Chinese picture mode Medium Mode Path For example, for the pattern path For any variable z, Figure 5 The matching nodes mapped to the variable z in the initial matching subgraph are as follows Figure 6 As shown, see Figure 6 , for any variable z, is the set of all matching nodes in the initial matching subgraph that can be mapped to the variable z, Upper bounds for variable scores, e.g., for pattern paths Medium Variable , the initial matching subgraph is mapped to this variable The matching nodes are and ,in, The node score is 0.6, The node score is 0.9, then it can be determined that the variable The upper bound of the variable score is =0.9; like Figure 7 , which is a schematic diagram of an influence score provided in Example 2 of the present application; In the process of determining the target node, the above pattern path For example, for the pattern path For any variable z in , the impact score of removing any matching node mapped to the variable on the variable score upper bound is as follows: Figure 7 As shown, see Figure 7 , for any variable z, for Any matching node in To score the impact, for example, for the pattern path Medium Variable ,variable The upper bound of the variable score (such as Figure 6 ) is 0.9, and the initial matching subgraph is mapped to this variable The matching nodes are and ; If removed , remove After mapping to variables The only matching nodes left are , then the variable The upper bound of the variable score will become 0.9, then according to the variables before and after removal The upper bound of the variable score can be calculated to remove For variables The variable score upper bound influences the score is 0.9-0.9=0; accordingly, if we remove , remove After mapping to variables The only matching nodes left are , then the variable The upper bound of the variable score will become 0.6, then according to the variables before and after removal The upper bound of the variable score can be calculated to remove For variables The variable score upper bound influences the score is 0.9-0.6=0.3; Figure 7 As shown, the maximum impact score can be determined is 0.3, then it can be determined is the target node.
[0036] Step S204, for any candidate evidence, determine the evidence score of the candidate evidence based on the rule score representing the importance of the interpretation rule and the path score representing the importance of the corresponding matching path, traverse all candidate evidence to obtain the evidence score of each candidate evidence, and obtain the target evidence based on the evidence scores of all candidate evidence and the initial score threshold.
[0037] The rule score may refer to a score representing the importance of the interpretation rule, the path score may refer to a score representing the importance of the matching path (matching result), the evidence score may refer to a score representing the importance of the evidence, and the target evidence may refer to candidate evidence whose evidence score reaches an initial score threshold.
[0038] Specifically, for any candidate evidence, the evidence score of the candidate evidence is determined based on the rule score that represents the importance of the interpretation rule and the path score that represents the importance of the corresponding matching path. All candidate evidences are traversed to obtain the evidence score of each candidate evidence, and the candidate evidence whose evidence score reaches the initial score threshold is determined as the target evidence.
[0039] Step S205, write the target evidence into the heap structure, use the lowest evidence score in the heap structure to update the initial score threshold, obtain the updated score threshold, remove the target node and the child nodes whose parent nodes are the target node from the initial matching subgraph, obtain the updated matching subgraph, use the updated matching subgraph to update the initial rule score upper bound, and obtain the updated rule score upper bound.
[0040] In this embodiment, the heap structure is used to store target evidence that meets the conditions. The target evidence that meets the conditions may refer to target evidence with a high priority (higher evidence score).
[0041] Specifically, the target evidence is written into the heap structure. After the writing is completed, the lowest evidence score in the heap structure is used to update and increase the initial score threshold, thereby obtaining an updated score threshold, so that the updated score threshold is applied as the new initial score threshold in step S206, and if the update rule score upper bound reaches the update score threshold, the updated matching subgraph is used as the initial matching subgraph, and the step of determining a target node from the initial matching subgraph is returned to be re-executed, or if the update rule score upper bound does not reach the update score threshold, when traversing all interpretation rules, it is ensured that the initial score threshold is continuously increasing, so that according to the continuously increasing initial score threshold, interpretation rules with higher initial rule score upper bounds can be screened and obtained, and evidence involving these interpretation rules with higher initial rule score upper bounds will contribute to the target evidence with high priority (higher evidence score) in the heap structure.
[0042] In the process of obtaining the updated matching subgraph, the target node and the child nodes whose parent nodes are the target node are removed from the initial matching subgraph. For example, Figure 5 The initial matching subgraph shown, and Figure 7 The target node shown , in the process of getting the updated matching subgraph, see Figure 5 , if you remove ,because yes The only parent node of At the same time, it is also necessary to It is also removed together to obtain the updated matching subgraph.
[0043] In the process of obtaining the upper bound of the update rule score, the contents in step S801 to step S803 may be referred to, and the upper bound of the update rule score may be obtained according to the updated matching subgraph.
[0044] Step S206, using the update score threshold as the initial score threshold, if the update rule score upper bound reaches the update score threshold, then using the update matching subgraph as the initial matching subgraph, returning to execute to determine a target node from the initial matching subgraph, and / or, if the update rule score upper bound does not reach the update score threshold, then traversing all interpretation rules and returning all target evidence in the heap structure.
[0045] Specifically, the update score threshold is applied as a new initial score threshold. If the update rule score upper bound reaches the update score threshold, the update matching subgraph is applied as a new initial matching subgraph, and the step of determining a target node from the initial matching subgraph in the above step S203 is returned and re-executed. The interpretation rule is continued to be processed until the update rule score upper bound of the interpretation rule does not reach the update score threshold, and the processing of the interpretation rule is stopped. The interpretation rules that have not been processed in the N interpretation rules are traversed, and the step of using any interpretation rule in the above step S202 to extract the initial matching subgraph corresponding to the user-item pair from the graph data is executed until all N interpretation rules have been processed (the initial rule score upper bounds of the N interpretation rules do not reach the initial score threshold, or the update rule upper bound does not reach the update score threshold), and all target evidence in the heap structure is returned; If the upper bound of the update rule score does not reach the update score threshold, stop processing the interpretation rule directly, traverse the unprocessed interpretation rules among the N interpretation rules, execute the above step S202 to use any interpretation rule to extract the initial matching subgraph corresponding to the user-item pair from the graph data, until all N interpretation rules have been processed and all target evidence in the heap structure are returned.
[0046] In the embodiment of the present application, the recommendation results of the recommendation model for recommending items to users in the graph data are explained through interpretation rules. On the basis of the subgraph, the topological structure and node features that have a decisive influence on the recommendation results are extracted to form target evidence, and the interpretation method is optimized, which further helps users understand which features play a key role in the model decision, thereby enhancing users' trust in the model results, improving the effectiveness (fidelity and sparsity) and reliability of the interpretation, and based on the rule score upper bound and the initial score threshold, it is ensured that high-priority (higher evidence score) target evidence that has a decisive influence on the recommendation results can be efficiently screened out, reducing the computational complexity and improving the efficiency of the algorithm. The target evidence with a high priority (higher evidence score) is dynamically updated and stored through the maintained heap structure, so that users can quickly obtain the most explanatory and influential evidence, thereby improving the user experience.
[0047] See also Figure 8 , is a flowchart of a rule-based recommendation model interpretation method provided in Example 3 of the present application. Figure 8 As shown, in the above step S202, determining the upper bound of the initial rule score of the interpretation rule according to the node score of the matching node in the initial matching subgraph may include the following steps: Step S801, determining the variables in the interpretation rule.
[0048] Step S802: for any variable, determine all matching nodes mapped to the variable in the initial matching subgraph, and determine the highest node score from the node scores of all matching nodes mapped to the variable as the variable score upper bound of the variable.
[0049] Step S803, obtaining an initial rule score upper bound of the interpretation rule according to the rule score representing the importance of the interpretation rule and the variable score upper bounds of all variables in the interpretation rule.
[0050] In this embodiment, the variable may refer to a predicate variable in the premise of the interpretation rule, the interpretation rule includes at least one variable, and the variable score upper bound may refer to the highest node score among all matching nodes mapped to the variable in the initial matching subgraph.
[0051] Specifically, the rules of interpretation The calculation formula for the upper bound of the initial rule score can be expressed as: in, is the upper bound of the initial rule score, Score the rule, is the upper bound of the variable score; That is, in the process of obtaining the initial rule score upper bound of the interpretation rule, for any variable in the interpretation rule, all matching nodes mapped to the variable are determined in the initial matching subgraph, and the highest node score is determined from the node scores of all matching nodes mapped to the variable as the variable score upper bound of the variable; the variable score upper bounds of all variables in the interpretation rule are added together to obtain the sum result, and the sum result and the rule score representing the importance of the interpretation rule are multiplied to obtain the initial rule score upper bound of the interpretation rule.
[0052] Optionally, after the variable score upper bounds of all variables in the interpretation rule are calculated, the variable score upper bounds of all variables in the interpretation rule are added together, and the sum result is directly used as the initial rule score upper bound of the interpretation rule.
[0053] In the embodiment of the present application, an initial rule score upper bound of the interpretation rule is obtained according to the rule score that characterizes the importance of the interpretation rule and the upper bound of the variable scores of all variables in the interpretation rule. Based on the initial rule score upper bound and the initial score threshold, evidence with low priority (lower evidence score) can be filtered out early, and finally evidence with high priority (higher evidence score) can be obtained, thereby reducing the computational complexity and improving the efficiency of the algorithm.
[0054] See also Fig. 9 , is a flowchart of a rule-based recommendation model interpretation method provided in Example 4 of the present application. Fig. 9As shown, in the above step S204, for any candidate evidence, determining the evidence score of the candidate evidence according to the rule score representing the importance of the interpretation rule and the path score representing the importance of the corresponding matching path may include the following steps: Step S901 , calculating a rule score representing the importance of the interpretation rule, and calculating a path score representing the importance of the corresponding matching path.
[0055] Step S902: Multiply the rule score and the path score to obtain the evidence score.
[0056] In this embodiment, the rule score is used to characterize the importance of the information expressed by the interpretation rule, and the path score is used to characterize the contribution of the matching path (matching result) to the recommendation model in the user-item pair prediction result.
[0057] Specifically, the calculation formula of the evidence score can be expressed as: in, Score the evidence. Score the rule, Score the path; Furthermore, for the form The calculation formula of the rule score can be expressed as: in, To explain the rules Chinese picture mode The pattern score, To explain the rules Prerequisites The condition score of For the pattern score , you can set the mode The number and length of paths in the formula are considered comprehensively to give priority to concise patterns; for conditional scores , the Gini Index (GINI) value can be used as a measure of the prerequisite How to classify the interpretation rule matching results (matching paths) into different categories based on the predictions of the recommendation model; Furthermore, the path score The calculation formula can be expressed as: in, Any matching node in the matching path Node score of For node scores , you can define the matching node It is obtained by comprehensively considering the actual value and degree of . Intuitively, nodes with larger degrees (or nodes with actual values closer to the optimal value, for example, nodes with higher scores are more popular) can be given higher scores. That is, in the process of obtaining the evidence score of the candidate evidence, first, the rule score representing the importance of the interpretation rule is calculated. The graph pattern corresponding to the interpretation rule and the premise of the interpretation rule are determined, the pattern score representing the importance of the graph pattern is calculated, and the condition score representing the importance of the premise is calculated, and the pattern score and the condition score are multiplied to obtain the rule score, and then the path score representing the importance of the corresponding matching path is calculated. The node scores of all matching nodes in the matching path are determined, and the node scores of all matching nodes in the matching path are added to obtain the path score representing the importance of the corresponding matching path. Finally, the rule score and the path score are multiplied to obtain the evidence score.
[0058] In the embodiment of the present application, an evidence score is calculated based on a rule score that characterizes the importance of the interpretation rule and a path score that characterizes the importance of the corresponding matching path. The importance of the rule and the importance of the subgraph and node features are comprehensively considered. The obtained evidence score can more accurately reflect the role of the target evidence in the process of the recommendation model obtaining the recommendation result, thereby improving the effectiveness and reliability of the explanation.
[0059] See also Fig.10 , is a flowchart of a rule-based recommendation model interpretation method provided in Example 5 of the present application, such as Fig.10 As shown, the explanation method of the recommendation model may further include the following steps: Step S1001, constructing a heap structure of size k; Writing the target evidence into the heap structure in the above step S205 includes: Step S1002, if the heap structure is not full, directly write the target evidence into the heap structure; Step S1003: if the heap structure is full, after removing the target evidence with the lowest evidence score from the heap structure, the target evidence is written into the heap structure.
[0060] In this embodiment, the heap structure is used to store target evidence that meets the conditions. The target evidence that meets the conditions may refer to the first k target evidences with high priority (higher evidence score), where k is an integer greater than zero.
[0061] If the heap structure is not full, the target evidence is directly written into the heap structure; if the heap structure is full, the target evidence with the lowest evidence score is removed from the heap structure and then written into the heap structure, so that after traversing all interpretation rules, the top k target evidences with high priority (higher evidence score) in the heap structure are returned.
[0062] For example, in order to obtain the top k target evidences with high priority (higher evidence score), a heap structure of size k is constructed, and the overall process of the explanation method of the recommendation model can be: 1) For any interpretation rule, calculate the upper bound of its initial rule score. If the upper bound of the initial rule score does not reach the initial score threshold, stop processing the interpretation rule based on the pruning strategy; 2) If the upper bound of the initial rule score reaches the initial score threshold, continue to process the interpretation rule, determine a target node and all matching paths containing the target node from the initial matching subgraph, and form a candidate evidence with each matching path and the interpretation rule; 3) Calculate the evidence score of each candidate evidence, and determine the candidate evidence whose evidence score reaches the initial score threshold as the target evidence; 4) Write the target evidence into the heap structure. During the writing process, if the heap structure is not full, directly write the target evidence into the heap structure. If the heap structure is full, write the target evidence into the heap structure after removing the target evidence with the lowest evidence score from the heap structure. After writing is completed, use the lowest evidence in the heap structure to obtain Update the initial score threshold to obtain the updated score threshold, remove the target node and the child nodes whose parent nodes are the target node from the initial matching subgraph to obtain the updated matching subgraph, use the updated matching subgraph to update the initial rule score upper bound to obtain the updated rule score upper bound; 5) Use the updated score threshold as the new initial score threshold for application. If the updated rule score upper bound reaches the updated score threshold, continue to process the interpretation rule, apply the updated matching subgraph as the new initial matching subgraph, and return to execute the above 2), 3), and 4) until the updated rule score upper bound does not reach the updated score threshold. Based on the pruning strategy, stop processing the interpretation rule, traverse all unprocessed interpretation rules, and execute the above 1), 2), 3), 4), and 5) until all interpretation rules have been processed, and return the top k target evidences with high priority (higher evidence scores) in the heap structure.
[0063] In an embodiment of the present application, a heap structure of size k is constructed. After obtaining the target evidence, if the heap structure is not full, the target evidence is directly written into the heap structure. If the heap structure is full, the target evidence is written into the heap structure after the target evidence with the lowest evidence score is removed from the heap structure. This allows the heap structure to always dynamically update and store the k target evidences that currently have a high priority (higher evidence score), so that users can quickly obtain the most explanatory and influential evidence, thereby improving the user experience.
[0064] Corresponding to the rule-based recommendation model interpretation method of the above embodiment, Fig.11 The structure block diagram of the rule-based recommendation model interpretation device provided in the sixth embodiment of the present application is shown. The recommendation model interpretation device is applied to Figure 1For ease of description, only the parts related to the embodiment of the present application are shown.
[0065] See also Fig.11 , the recommendation model explanation device comprises: The acquisition module 1101 is used to acquire N interpretation rules, graph data, and user-item pairs to be interpreted in the graph data, wherein the user-item pairs are node pairs representing the mapping relationship between users and items obtained by recommending items to users in the graph data based on a recommendation model, and N is an integer greater than zero; A rule score determination module 1102 is used to extract an initial matching subgraph corresponding to the user-item pair from the graph data using any interpretation rule, and determine an upper bound of an initial rule score of the interpretation rule according to node scores of matching nodes in the initial matching subgraph; A rule score judgment module 1103 is used to determine a target node and all matching paths including the target node from the initial matching subgraph when the initial rule score upper bound reaches the initial score threshold, and form a candidate evidence with each matching path and the interpretation rule; The target evidence judgment module 1104 is used to determine the evidence score of any candidate evidence according to the rule score representing the importance of the interpretation rule and the path score representing the importance of the corresponding matching path, traverse all candidate evidences to obtain the evidence score of each candidate evidence, and obtain the target evidence according to the evidence scores of all candidate evidences and the initial score threshold; An updating module 1105 is used to write the target evidence into a heap structure, update the initial score threshold using the lowest evidence score in the heap structure to obtain an updated score threshold, remove the target node and the child nodes whose parent nodes are the target node from the initial matching subgraph to obtain an updated matching subgraph, and use the updated matching subgraph to update the initial rule score upper bound to obtain an updated rule score upper bound. The heap structure is used to store target evidence that meets the conditions; The loop module 1106 is used to use the update score threshold as the initial score threshold, and if the update rule score upper bound reaches the update score threshold, use the update matching subgraph as the initial matching subgraph, and return to execute the determination of a target node from the initial matching subgraph, and / or, if the update rule score upper bound does not reach the update score threshold, traverse all interpretation rules and return all target evidence in the heap structure.
[0066] Optionally, the rule score determination module 1102 includes: a variable determination unit, configured to determine a variable in the interpretation rule, wherein the interpretation rule includes at least one variable; a variable score determination unit, configured to determine, for any variable, all matching nodes mapped to the variable in the initial matching subgraph, and determine the highest node score from the node scores of all matching nodes mapped to the variable as the variable score upper bound of the variable; The rule score calculation unit is used to obtain the initial rule score upper bound of the interpretation rule according to the rule score representing the importance of the interpretation rule and the variable score upper bounds of all variables in the interpretation rule.
[0067] Optionally, the rule score determination module 1103 includes: An influence score calculation unit, used for calculating, for any variable in the interpretation rule, an influence score of removing any matching node mapped to the variable on the variable score upper bound of the variable; The target node determination unit is used to determine the matching node with the largest influence score as the target node.
[0068] Optionally, the target evidence judgment module 1104 includes: An importance calculation unit, used to calculate a rule score representing the importance of the interpretation rule, and to calculate a path score representing the importance of the corresponding matching path; The first multiplication unit is used to perform a multiplication operation on the rule score and the path score to obtain the evidence score.
[0069] Optionally, the importance calculation unit includes: A first condition determination subunit, used to determine the graph pattern corresponding to the interpretation rule and the precondition of the interpretation rule; A condition calculation subunit, used to calculate a pattern score representing the importance of the graph pattern, and to calculate a condition score representing the importance of the prerequisite; The second multiplication subunit is used to perform a multiplication operation on the pattern score and the condition score to obtain a rule score representing the importance of the interpretation rule.
[0070] Optionally, the importance calculation unit includes: A second condition determination subunit, used to determine the node scores of all matching nodes in the matching path; The summing subunit is used to sum the node scores of all matching nodes in the matching path to obtain a path score representing the importance of the corresponding matching path.
[0071] Optionally, the recommended model interpretation device further includes: A construction module is used to construct a heap structure of size k, where k is an integer greater than zero; The update module 1105 includes: A first writing unit, configured to directly write the target evidence into the heap structure if the heap structure is not full; The second writing unit is used for writing the target evidence into the heap structure after removing the target evidence with the lowest evidence score from the heap structure if the heap structure is full.
[0072] It should be noted that the information interaction, execution process and other contents between the above-mentioned modules are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.
[0073] Fig.12 This is a schematic diagram of the structure of a computer device provided in Example 7 of this application. Fig.12 As shown, the computer device of this embodiment includes: at least one processor ( Fig.12 Only one is shown in the figure), a memory, and a computer program stored in the memory and executable on at least one processor, which implements the steps of any of the above-mentioned rule-based recommendation model interpretation method embodiments when the processor executes the computer program.
[0074] The computer device may include, but is not limited to, a processor and a memory. Those skilled in the art will appreciate that Fig.12 This is merely an example of a computer device and does not constitute a limitation on the computer device. The computer device may include more or fewer components than those shown in the figure, or a combination of certain components, or different components. For example, it may also include a network interface, a display screen, and an input device.
[0075] The processor may be a CPU, or other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.
[0076] The memory includes a readable storage medium, an internal memory, etc., wherein the internal memory may be the memory of a computer device, and the internal memory provides an environment for the operation of the operating system and computer-readable instructions in the readable storage medium. The readable storage medium may be a hard disk of a computer device, and in other embodiments, it may also be an external storage device of the computer device, for example, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the computer device. Further, the memory may also include both an internal storage unit of the computer device and an external storage device. The memory is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of a computer program, etc. The memory may also be used to temporarily store data that has been output or is to be output.
[0077] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned device can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here. If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned method embodiment can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device capable of carrying computer program code, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.
[0078] The present application implements all or part of the processes in the above-mentioned embodiment method, and may also be completed through a computer program product. When the computer program product runs on a computer device, the steps in the above-mentioned method embodiment can be implemented when the computer device executes it.
[0079] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0080] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0081] In the embodiments provided in the present application, it should be understood that the disclosed devices / computer equipment and methods can be implemented in other ways. For example, the device / computer equipment embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.
[0082] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0083] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.
Claims
1. A rule-based recommendation model interpretation method, characterized in that: The recommendation model interpretation method comprises: Obtaining N interpretation rules, graph data, and user-item pairs to be interpreted in the graph data, wherein the user-item pairs are node pairs representing a mapping relationship between users and items obtained by recommending items to users in the graph data based on a recommendation model, and N is an integer greater than zero; Using any interpretation rule, extracting an initial matching subgraph corresponding to the user-item pair from the graph data, and determining an initial rule score upper bound of the interpretation rule according to node scores of matching nodes in the initial matching subgraph; When the upper bound of the initial rule score reaches the initial score threshold, a target node and all matching paths including the target node are determined from the initial matching subgraph, and each matching path is combined with the interpretation rule to form a candidate evidence; For any candidate evidence, determine the evidence score of the candidate evidence according to the rule score representing the importance of the interpretation rule and the path score representing the importance of the corresponding matching path, traverse all candidate evidences to obtain the evidence score of each candidate evidence, and obtain the target evidence according to the evidence scores of all candidate evidences and the initial score threshold; The target evidence is written into a heap structure, the lowest evidence score in the heap structure is used to update the initial score threshold to obtain an updated score threshold, the target node and the child nodes whose parent nodes are the target node are removed from the initial matching subgraph to obtain an updated matching subgraph, the initial rule score upper bound is updated using the updated matching subgraph to obtain an updated rule score upper bound, and the heap structure is used to store target evidence that meets the conditions; The update score threshold is used as the initial score threshold. If the update rule score upper bound reaches the update score threshold, the update matching subgraph is used as the initial matching subgraph, and the determination of a target node from the initial matching subgraph is returned to be executed, and / or, if the update rule score upper bound does not reach the update score threshold, all interpretation rules are traversed to return all target evidences in the heap structure.
2. The recommendation model interpretation method according to claim 1, characterized in that: The step of determining the upper bound of the initial rule score of the interpretation rule according to the node score of the matching node in the initial matching subgraph includes: determining variables in the interpretation rule, the interpretation rule comprising at least one variable; For any variable, determine all matching nodes mapped to the variable in the initial matching subgraph, and determine the highest node score from the node scores of all matching nodes mapped to the variable as the variable score upper bound of the variable; An initial rule score upper bound of the interpretation rule is obtained according to a rule score representing the importance of the interpretation rule and an upper bound of variable scores of all variables in the interpretation rule.
3. The recommendation model interpretation method according to claim 2, characterized in that: The determining a target node from the initial matching subgraph includes: For any variable in the interpretation rule, calculate the impact score of removing any matching node mapped to the variable on the variable score upper bound of the variable; The matching node with the largest influence score is determined as the target node.
4. The recommendation model interpretation method according to claim 1, characterized in that: For any candidate evidence, determining the evidence score of the candidate evidence according to the rule score representing the importance of the interpretation rule and the path score representing the importance of the corresponding matching path includes: Calculating a rule score representing the importance of the interpretation rule, and calculating a path score representing the importance of the corresponding matching path; The rule score and the path score are multiplied to obtain the evidence score.
5. The recommendation model interpretation method according to claim 4, characterized in that: The calculating of the rule score representing the importance of the interpretation rule comprises: Determining a graph pattern corresponding to the interpretation rule and a precondition of the interpretation rule; Calculating a pattern score representing the importance of the graph pattern, and calculating a condition score representing the importance of the prerequisite condition; The pattern score and the condition score are multiplied to obtain a rule score representing the importance of the interpretation rule.
6. The recommendation model interpretation method according to claim 4, characterized in that: The calculating of the path score representing the importance of the corresponding matching path includes: Determining node scores of all matching nodes in the matching path; The node scores of all matching nodes in the matching path are summed up to obtain a path score representing the importance of the corresponding matching path.
7. The recommendation model interpretation method according to claim 1, characterized in that: The recommendation model interpretation method further includes: Construct a heap structure of size k, where k is an integer greater than zero; The step of writing the target evidence into the heap structure comprises: If the heap structure is not full, directly writing the target evidence into the heap structure; If the heap structure is full, the target evidence is written into the heap structure after the target evidence with the lowest evidence score is removed from the heap structure.
8. A rule-based recommendation model interpretation device, characterized in that: The recommendation model interpretation device comprises: an acquisition module, configured to acquire N interpretation rules, graph data, and user-item pairs to be interpreted in the graph data, wherein the user-item pairs are node pairs representing a mapping relationship between users and items obtained by recommending items to users in the graph data based on a recommendation model, and N is an integer greater than zero; a rule score determination module, configured to extract an initial matching subgraph corresponding to the user-item pair from the graph data using any interpretation rule, and determine an upper bound of an initial rule score of the interpretation rule according to node scores of matching nodes in the initial matching subgraph; A rule score judgment module, used for determining a target node and all matching paths including the target node from the initial matching subgraph when the upper bound of the initial rule score reaches the initial score threshold, and forming a candidate evidence with each matching path and the interpretation rule; A target evidence judgment module is used to determine the evidence score of any candidate evidence according to the rule score representing the importance of the interpretation rule and the path score representing the importance of the corresponding matching path, traverse all candidate evidences to obtain the evidence score of each candidate evidence, and obtain the target evidence according to the evidence scores of all candidate evidences and the initial score threshold; An updating module is used to write the target evidence into a heap structure, update the initial score threshold using the lowest evidence score in the heap structure to obtain an updated score threshold, remove the target node and the child nodes whose parent nodes are the target node from the initial matching subgraph to obtain an updated matching subgraph, and use the updated matching subgraph to update the initial rule score upper bound to obtain an updated rule score upper bound. The heap structure is used to store target evidence that meets the conditions; A loop module is used to use the update score threshold as the initial score threshold, and if the update rule score upper bound reaches the update score threshold, use the update matching subgraph as the initial matching subgraph, return to execute the determination of a target node from the initial matching subgraph, and / or, if the update rule score upper bound does not reach the update score threshold, traverse all interpretation rules and return all target evidence in the heap structure.
9. A computer device, characterized in that: The computer device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the recommendation model interpretation method according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the recommendation model interpretation method according to any one of claims 1 to 7 is implemented.