A method for product search with query-dependent multi-faceted explainability
By using a query-aware graph convolutional network, combined with a query-aware graph convolutional sorter and a multi-path inferencer, the problems of query mismatch and one-sided interpretation in existing technologies are solved, enabling product search with multiple interpretations and improving retrieval performance and user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING NORMAL UNIVERSITY
- Filing Date
- 2024-01-18
- Publication Date
- 2026-06-26
AI Technical Summary
Existing product search methods neglect the importance of queries during the interpretation and generation process, resulting in a mismatch between the path and the user's query. Furthermore, most models only provide one-sided interpretations, making it difficult to meet the diverse search intentions of users.
A query-aware graph convolutional network is adopted, including a query-aware graph convolutional sorter and a query-aware multi-path inferencer. The graph convolutional network models user and product representations, and uses knowledge graphs to generate multifaceted explanations, combining query relevance and user search intent.
It significantly improves retrieval performance, generates more accurate and multifaceted explanations, meets diverse user search intents, and enhances user experience and trust in the platform.
Smart Images
Figure CN117911110B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a product search method that provides query-related, multi-faceted interpretability. Background Technology
[0002] Among the key tools of modern e-commerce platforms, product search plays a crucial role in helping users explore and purchase goods from a wide range of choices. In a real-world product search scenario, users describe their needs to the search engine—that is, they enter a query. The search engine then responds with a complete list of products relevant to the given query, which are displayed on the results page. By browsing these results pages, users can easily select and purchase products that interest them. This seamless process enhances the overall online shopping experience, making it more efficient and convenient for users to find and buy the goods they want.
[0003] Traditional product search methods primarily focus on matching queries with various aspects of product information (such as brand, category, and context), often providing users with non-personalized search results. In real-world product search scenarios, user purchasing behavior is influenced by personal interests and preferences. This means users have different preferences for specific product attributes (such as color, brand, and price), which affects their decision-making process. To meet diverse user needs, many personalized product search models have been proposed, utilizing users' historical logs (including their purchase history and product / user reviews) to capture and model their personalized preferences. While personalized methods enhance product search retrieval performance by providing user-centric search results, displaying only relevant products can sometimes confuse users. Users are unsure of the relationship between the displayed products and their submitted queries, leading to user dissatisfaction and distrust of the platform. Given the crucial role of explanation in improving user satisfaction and trust in e-commerce platforms, many product search methods attempt to leverage knowledge graphs to generate potential explanations for search results, making the product search process more transparent to users.
[0004] While previous interpretable product search efforts have improved retrieval efficiency to some extent, these methods still have inherent limitations. First, many models neglect the importance of treating the current query as a prerequisite during the explanation generation process; therefore, they may generate explanations that are irrelevant to the query and between the user and the item. The mismatch between path-based explanations and the user's current query can lead to further user confusion. Second, existing models are limited to providing one-sided explanations, offering only a single reasoning path within the knowledge graph to interpret the search results. Although some strategies can be employed to generate multi-faceted explanations in the form of multiple paths, these paths are often modeled independently, which makes it difficult to satisfy diverse user search intents, as users typically make purchasing decisions by comparing multiple aspects of a product. To address these issues, we focus on query relevance in explanation generation, while providing multi-faceted explanations through multiple reasoning paths related to different user search intents. Summary of the Invention
[0005] To this end, the present invention first proposes a product search method with multi-faceted interpretability related to querying. The method involves inputting product search information, using a multi-path query-aware graph convolutional network for interpretable product search to perform the search query and return the product search results.
[0006] The interpretable product search multi-path query-aware graph convolutional network comprises two components: a query-aware graph convolutional sorter and a query-aware multi-path inferencer. The query-aware graph convolutional sorter uses graph convolutional networks to model user and product representations based on different knowledge relation domains in the knowledge graph. The query-aware multi-path inferencer is responsible for exploring query-specific multi-paths from the knowledge graph to satisfy the user's search intent. The two components share basic parameters and are co-trained to form a complete network.
[0007] The knowledge graph upon which the graph convolutional network in the query-aware graph convolutional sorter is based is defined as: composed of multiple triples, i.e. , in which Connected to the head entity and tail body , in and These are the relationship set and entity set of the Key-Value Group (KG), with five entity types: user, product, word, category, and brand, and eight relationships between entities: Is_brand, Is_category, search & Purchase, Write, Mentioned, Also_bought, Also_viewed, and Bought_together, which together construct the user-product KG. A query consists of n words. The query result is calculated as follows:
[0008]
[0009] and Represents training parameters, On behalf of users, Representative products , and They represent the relationship and belonging to a certain category, respectively. Related entities, given a user And the current query The model will return a list of products based on their relevance to the user's search intent. At the same time, it generates multiple interpretations in the form of multiple paths, such as ,in It is the path between products that satisfy a user's search intent for a specific aspect.
[0010] The query-aware graph convolutional sorter includes representation modeling and representation learning modules. The representation modeling module builds a query-aware convolution to model user and product representations based on multiple knowledge relationship domains. The representation learning module considers dynamic and static relationships respectively to optimize and learn more accurate user and product representations.
[0011] The representation modeling module designs a query-aware hierarchical graph convolution on the multi-hop neighbors of entities with different knowledge relation domains. For each layer of convolution, it considers modeling the entity representation in a specific knowledge relation domain based on the current query, and integrates entity representations from different knowledge relation domains based on the current query. Specifically, for entities... , This is the set of neighbors of the entity across different knowledge relation domains. Then, for a specific relation... ,That The hierarchical graph convolution for modeling layer-specific knowledge relation entity representations is as follows:
[0012]
[0013]
[0014] in Represents the dot product of digital vectors. It is the user in the relationship Neighboring entities in the middle, Represented as It is a physical entity The original representation, Indicates neighbors In relation The weights in the model represent the importance of the neighborhood in the current query. The weight function is defined. :
[0015] After establishing entity representations for a specific knowledge domain, a query-related attention mechanism is used to aggregate different relation-specific entity representations into a single entity representation. The aggregation process is as follows:
[0016]
[0017]
[0018] Representing relations In entity The weights in the adjacent all relations are obtained after... The entity representation of the jump is combined with the original entity representation and passed through the MLP layer to obtain the final entity representation, as shown below:
[0019]
[0020] Will and As a user and goods Representation obtained from query-aware hierarchical graph convolution.
[0021] The representation learning module uses TransE to define a scoring function to measure the plausibility of triples observed in the KG, letting Given a triple, and Observed The probability is:
[0022]
[0023] in In the relationship Given the set of all possible end entities, rewrite the user-query-product triple in the equation based on the characteristics of dynamic relationship search and purchasing. The probability of:
[0024]
[0025]
[0026] in Indicates specific The lower-level product collection, and This represents the maximum number of convolutional layers for users and items. The model parameters are directly optimized by maximizing the log-likelihood of the observed triples for all relationships, as shown below:
[0027]
[0028]
[0029] in It is the size of the negative samples. noise distribution The negative samples below, and This represents all observed positive samples in the dataset. This is the sigmoid function.
[0030] The query-aware multipath inferencer includes: process one, a process of heuristically extracting and selecting query-specific path examples, and process two, a process of modeling the user's search intent and generating multifaceted explanations guided by these examples.
[0031] The process includes three steps:
[0032] Step 1: Search for the path between the user and the product based on the predefined meta-path. First, extract the meta-path as a predefined set of paths through random walk, and keep the path to the product purchased by the user as candidate examples.
[0033] Step 2: Path selection for a specific query. Directly calculate the semantic relevance between the query in the path demonstration and the current query to determine the degree of relevance between the demonstration and the user's current search intent:
[0034]
[0035] It is a query along a path. The computation is performed using the original embeddings from the pre-trained BERT model; then, a set of path examples whose relevance to the current query is greater than a specific threshold is selected and retained as query-relevant paths, denoted as . And use it as a supervisory signal to guide the multi-path inferencer to generate multifaceted explanations;
[0036] Step 3: Establish positive and negative path signals, and use the query-related path examples found in Step 2. As a training triplet The strong positive signal is then used to replace the last product entity in each path with a negative example, thus obtaining the set of negative example paths. , as a strong negative signal of the triple.
[0037] The second process specifically involves: using path examples as monitoring signals to infer multipaths, for query-specific triples. The random sampling path serves as a multifaceted explanation for the products retrieved under the current query; that is, a long short-term memory network is first used. The paths are encoded, these path representations are then integrated, and their consistency with the user's current search intent is measured. The probability of matching the user's current search intent is calculated using multiple interpretations, as follows:
[0038]
[0039] During the training phase, negative samples with multiple interpretations were constructed for each triple. A multifaceted explanation of positive samples, using items that the user has never purchased under the current query. The endpoint entities of the path are replaced to obtain new negative sample paths. During training, this part of the problem is treated as a binary classification problem, and the cross-entropy loss is used to calculate the loss in the inference engine. The inference loss is:
[0040] .
[0041] The training process specifically involves: jointly optimizing the sorter and the multi-path inference engine, with a total loss of: ,in It is a hyperparameter that controls the importance of two components during training;
[0042] During the inference phase, the following steps are used for each triplet. Generate multifaceted explanations: First, using user... As the starting node, the paths to the goods are sampled, and their meta-paths belong to a predefined set of candidate multipaths. Then, through... Encode each candidate path and calculate its matching probability with the triple:
[0043]
[0044] Finally, these candidate paths are sorted in descending order of matching scores, and the paths with scores higher than a set threshold are selected as the final multivariate interpretation.
[0045] The technical effects to be achieved by this invention are as follows:
[0046] (1) A product search model with multi-faceted interpretability related to query is proposed to significantly improve retrieval performance and generate better explanations for search results.
[0047] (2) A query-aware graph convolutional sorting algorithm was used to obtain fine-grained representations of users and products by aggregating multi-hop neighborhoods with hierarchical query-aware attention.
[0048] (3) A query-aware multi-path inferencer was designed to explore multiple paths on the knowledge graph. It not only models the user’s multi-faceted preferences to obtain better retrieval performance, but also generates multi-faceted explanations. Attached Figure Description
[0049] Figure 1 A diagram of a multi-path query-aware convolutional network architecture that can explain product search. Detailed Implementation
[0050] The following are preferred embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.
[0051] This invention proposes a query-related, multi-faceted interpretability-based product search method. It is implemented through an interpretable product search model, namely, a multi-path query-aware graph convolutional network (QGCNM) for interpretable product search.
[0052] This network consists of a query-aware graph convolutional layer and a query-aware multipath inferencer. The ranking layer learns multifaceted user search intent and item representations through query-aware hierarchical convolutions based on the current query. The inferencer models user search intent using query-specific path demonstrations and explores multifaceted interpretations of search results. The two components share some fundamental parameters, such as jointly trained entity and relation representations. This joint training facilitates interaction between the two components, enabling the model to learn structural information present in the knowledge graph and develop reasoning capabilities. Empirical experiments demonstrate that our model significantly improves retrieval performance compared to baseline models and is able to generate multifaceted interpretations.
[0053] Knowledge graph construction:
[0054] Table 1 Relationship between User and Product Images
[0055]
[0056] A typical knowledge graph (KG) is a multi-relation directed graph, which consists of many triples, i.e. , in which Connected to the head entity and tail body , in and These are the relation set and entity set of the KG, respectively. Taking into full account the information in the dataset and the definitions from previous work, we selected five entity types (user, product, word, category, and brand) and eight relationships between entities (Is_brand, Is_category, search & Purchase, Write, Mentioned, Also_bought, Also_viewed, Bought_together) to construct the user-product KG, details of which are shown in Table 1. Search and purchase are considered dynamic relationships because their semantics are determined by the dynamic search content. For a query, assuming it consists of n words, The query result is calculated as follows:
[0057]
[0058] and Representing training parameters, in a KG-based interpretable product search method, the path connecting the user and the product is considered an interpretation of the retrieved product. For example, the path... This can represent the user Products found The explanation, and They represent the relationship and belonging to a certain category, respectively. Related entities. Therefore, given a user And the current query The model will return a list of products based on their relevance to the user's search intent. At the same time, it generates multiple interpretations in the form of multiple paths, such as ,in It is the path between products that satisfy a user's search intent for a specific aspect; that is, the path that explains why when Submit query Product search engine returns .
[0059] Query-aware graph convolutional sorter:
[0060] In order to make full use of the structural information of the knowledge graph to better learn user representations (i.e., multi-faceted preferences) and product representations from multiple knowledge relation domains, we designed a graph convolution-based ranking module that considers the current user query. This module mainly includes two aspects: (1) representation modeling, which builds a query-aware convolution to model user and product representations based on multiple knowledge relation domains; and (2) representation learning, which considers dynamic and static relations respectively to optimize and learn more accurate user and product representations.
[0061] Representation Modeling: To comprehensively model user representations, it is necessary to mine the distribution of user search intent from different aspects of the current query. Therefore, we designed a query-aware hierarchical graph convolution based on the multi-hop neighbors of entities with different knowledge relation domains. For each layer of convolution, we consider (1) modeling entity representations in a specific knowledge relation domain based on the current query, and (2) integrating entity representations from different knowledge relation domains based on the current query. For example, This is the set of neighbors of the entity across different knowledge relation domains. Then, for a specific relation... ,That The hierarchical graph convolution for modeling layer-specific knowledge relation entity representations is as follows:
[0062]
[0063]
[0064] in Represents the dot product of digital vectors. It is the user in the relationship Neighboring entities in Represented as It is a physical entity The original representation, Indicates neighbors In relation The weights in the model. Modeling the importance of neighborhoods in the current query. We defined the weight function. :
[0065]
[0066] After establishing entity representations for a specific knowledge domain, we employ a query-related attention mechanism to aggregate different relation-specific entity representations into a single entity representation. The aggregation process is as follows:
[0067]
[0068]
[0069] Representing relations In entity The weights in the adjacent all relations are obtained after... We combine the skipped entity representation with the original entity representation and pass it through an MLP layer to obtain the final entity representation, as shown below:
[0070]
[0071] To fully express users' diverse search intentions and the relevant product representations, we will and As a user and goods Representation obtained from query-aware hierarchical graph convolution.
[0072] Representation learning: We use TransE to define a scoring function to measure the plausibility of triples observed in the KG. Let Given a triple, and Observed The probability is:
[0073]
[0074] in In the relationship The set of all possible end entities. Based on the characteristics of dynamic relationship search and purchasing, we rewrote the user-query-product triple in the equation. The probability of:
[0075]
[0076]
[0077] in Indicates specific The lower-level product collection, and This represents the maximum number of convolutional layers for users and items. To reduce computational costs, we directly optimize the model parameters by maximizing the log-likelihood of observed triples for all relationships, as shown below:
[0078]
[0079]
[0080] in It is the size of the negative samples. noise distribution The negative samples below, and This represents all observed positive samples in the dataset. This is the sigmoid function.
[0081] Query-aware multipath inferencer:
[0082] To generate explanations that satisfy users’ complex search intents, we introduce a query-aware multi-path inferencer, which includes (1) heuristically extracting and selecting query-specific path examples, and (2) modeling the user’s search intent and generating multifaceted explanations guided by these examples.
[0083] Heuristic path extraction. The path example is a query-specific multi-hop path between a user and a product in the user-product KG. To demonstrate heuristic path extraction, there are three main steps:
[0084] Searching for paths between users and products based on predefined metapaths. A metapath is a sequence of entity and relation types, such as... Based on the user-product KG graph and the types of relationships and entities, we first extract meta-paths as paths in a predefined set through random walks, and retain paths leading to products purchased by users as candidate examples.
[0085] Path selection for a specific query. A path example is valid if it contains queries related to or identical to the current query, ensuring consistency between the interpretation and the user's current search intent. Based on the characteristics of the meta-path, we directly calculate the semantic relevance between the queries in the path demonstration and the current query to determine the degree of relevance between the demonstration and the user's current search intent. The specific details of the relevance calculation are as follows:
[0086]
[0087] It is a query along a path. We perform computation using the raw embeddings from the pre-trained BERT model. Then, we select and retain the set of path examples whose relevance to the current query is greater than a certain threshold as query-relevant paths, denoted as . This is used as a supervisory signal to guide the multi-path inferencer in generating multifaceted explanations.
[0088] Establish positive and negative path signals. Use the query-related path examples found in step (2). As a training triplet A strong positive signal. Then, replace the last product entity in each path with a negative example to obtain a set of negative example paths. , as a strong negative signal of the triple.
[0089] Multi-path explanation generation. To enable the multi-path inference engine to generate multifaceted explanations, we use path examples as supervision signals to infer multi-paths. Specifically, for query-specific triples... We randomly sample paths as multifaceted interpretations of the retrieved products under the current query. First, we encode the paths using a Long Short-Term Memory (LSTM) network to better capture the inference and sequence information of the paths. Second, we integrate these path representations and use them to measure their consistency with the user's current search intent. The probability of matching multifaceted interpretations with the user's current search intent is calculated as follows:
[0090]
[0091] To improve the robustness of the inference engine, we constructed negative examples with multifaceted explanations during the training phase. For each triplet... To provide a more comprehensive explanation of positive samples, we use items that the user has never purchased under the current query. This replaces the endpoint entities of the path to obtain new negative sample paths. During training, we treat this part of the problem as a binary classification problem and use cross-entropy loss to calculate the loss in the inference engine. The inference loss is:
[0092]
[0093] Training and Reasoning:
[0094] During the training phase, to ensure that the model's parameters not only capture ranking information but also model the interpretation, the sorter and multi-path inference will be jointly optimized:
[0095]
[0096] in It is a hyperparameter that controls the importance of two components during training.
[0097] During the inference phase, the following steps are used for each triplet. Generate multifaceted explanations (first, using user...) As the starting node, we sample the paths to the goods, and their meta-paths belong to a predefined set of candidate multipaths. Next, we... Encode each candidate path and calculate its matching probability with the triple:
[0098]
[0099] Finally, these candidate paths are sorted in descending order of matching scores, and the paths with scores higher than a set threshold are selected as the final multivariate interpretation.
[0100] Experimental results:
[0101] Table 2 Overall Model Performance
[0102]
[0103] To demonstrate the effectiveness of the QGCNM model in the product search domain, six benchmark models and four Amazon subsets were selected, and the model's performance on different datasets was compared. The overall performance is shown in the table below.
[0104] Based on the analysis of the summary table:
[0105] (1) The model outperforms other benchmark models because it can better uncover users' multi-interest search intent based on the current query through multi-faceted interpretations. Compared with the benchmark models, QGCNM achieves significant improvements in all metrics on each dataset, with QGCNM improving MAP and MRR by at least 13.27%, and NDCG@10 and NDCG@20 performance by at least 21.67% and 19.79%, respectively. Therefore, the key to improving retrieval performance is to consider the user's current query in user (product) representation modeling and multi-faceted interpretation generation. This indicates that the current query is important for improving retrieval performance and generating accurate interpretations.
[0106] (2) The model achieved the highest performance among interpretable models because the multi-path inferencer can enhance the user's search intent and explore multiple interpretations. In the model, the multi-path inferencer heuristically generates path examples based on the user's historical search behavior. On the one hand, these examples carry the user's historical search preferences and more implicit information about the user's current search intent for retrieval. On the other hand, they can provide supervised signals for the generation of multiple interpretations, which ensures that the interpretations are consistent with the user's search intent. Figure 1 The key to this.
[0107] (3) The model achieves better performance than other personalized product search methods because the sorter applies query-aware graph convolutions on the user's product KG to model user and product representations. Specifically, this invention develops a query-aware hierarchical convolution over multiple knowledge domains, which can obtain multifaceted search intents and product representations of users in these aspects for more accurate searches.
[0108] (4) KG-based models have better retrieval performance than sequence-based models. Sequence-based models like HEM and ZAM utilize the user's search sequence to model the user's search intent. In contrast, KG-based product search models, including SBG, DREM, DREM-HGM, and QGCNM, can fully utilize the structural information of the user's product KG to establish a matching relationship between the user and the product under the submitted query. This indicates that using KG-based structural information in product search scenarios can better determine the user's preferences and purchase intent, thereby improving retrieval performance.
Claims
1. A product search method for querying relevant multi-faceted interpretability, characterized in that: Input the product search information, use the multi-path query-aware graph convolutional network that can interpret product search to perform the search query and return the search results; The interpretable product search multi-path query-aware graph convolutional network includes two components: a query-aware graph convolutional sorter and a query-aware multi-path inferencer. The query-aware graph convolutional sorter uses graph convolutional networks to model user and product representations based on different knowledge relation domains in the knowledge graph. The query-aware multi-path inferencer is responsible for exploring query-specific multi-paths from the knowledge graph to meet the user's search intent. The two components share basic parameters and are co-trained to form a complete network. The query-aware graph convolutional sorter includes a representation modeling module and a representation learning module. The representation modeling module models user and product representations by constructing a query-aware convolution and based on multiple knowledge relationship domains. The representation learning module considers dynamic and static relationships respectively to optimize and learn more accurate user and product representations. The representation modeling module designs a query-aware hierarchical graph convolution on the multi-hop neighbors of entities with different knowledge relation domains. For each layer of convolution, it considers modeling the entity representation in a specific knowledge relation domain based on the current query, and integrates entity representations from different knowledge relation domains based on the current query. Specifically, for entities... , The set of neighbors of the entity across different knowledge relation domains, for a specific relation ,That The hierarchical graph convolution for modeling layer-specific knowledge relation entity representations is as follows: in Represents the dot product of digital vectors. It is the user in the relationship Neighboring entities in Represented as It is a physical entity The original representation, Indicates neighbors In relationship The weights in the model represent the importance of the neighborhood in the current query. The weight function is defined. : ; After establishing entity representations for a specific knowledge domain, a query-related attention mechanism is used to aggregate different relation-specific entity representations into a single entity representation. The aggregation process is as follows: Representing relations In entity The weights in the adjacent all relations are obtained after... The entity representation of the jump is combined with the original entity representation and passed through the MLP layer to obtain the final entity representation, as shown below: Will and As a user and goods Representation obtained from query-aware hierarchical graph convolution; The query-aware multipath inferencer includes: process one, a process of heuristically extracting and selecting query-specific path examples, and process two, a process of modeling the user's search intent and generating multifaceted explanations guided by these examples.
2. The product search method for querying relevant multi-faceted interpretability as described in claim 1, characterized in that: The knowledge graph upon which the graph convolutional network in the query-aware graph convolutional sorter is based is defined as: composed of multiple triples, i.e. , in Connected to the head entity and tail body , in and These are the relationship set and entity set of the Key-Value Group (KG), with five entity types: user, product, word, category, and brand, and eight relationships between entities: Is_brand, Is_category, search & Purchase, Write, Mentioned, Also_bought, Also_viewed, and Bought_together, which together construct the user-product KG. A query consists of n words. The query result is calculated as follows: and Represents training parameters, On behalf of users, Representative products , Representational relationship, given user And the current query The model will return a list of products based on their relevance to the user's search intent. At the same time, it generates multiple interpretations in the form of multiple paths, such as ,in It is the path between products that satisfy a user's search intent for a specific aspect.
3. The product search method for querying relevant multi-faceted interpretability as described in claim 2, characterized in that: The representation learning module uses TransE to define a scoring function to measure the plausibility of triples observed in the KG, letting Given a triple, and Observed The probability is: in In the relationship Given the set of all possible end entities, rewrite the user-query-product triple in the equation based on the characteristics of dynamic relationship search and purchasing. The probability of: in Indicates specific The lower-level product collection, and This represents the maximum number of convolutional layers for users and items. The model parameters are directly optimized by maximizing the log-likelihood of the observed triples for all relationships, as shown below: in It is the size of the negative samples. noise distribution The negative samples below, and This represents all observed positive samples in the dataset. This is the sigmoid function.
4. The product search method for querying relevant multi-faceted interpretability as described in claim 3, characterized in that: The process includes three steps: Step 1: Search for the path between the user and the product based on the predefined meta-path. First, extract the meta-path as a predefined set of paths through random walk, and keep the path to the product purchased by the user as candidate examples. Step 2: Path selection for a specific query. Directly calculate the semantic relevance between the query in the path demonstration and the current query to determine the degree of relevance between the demonstration and the user's current search intent: It is a query along a path. The computation is performed using the original embeddings from the pre-trained BERT model; then, a set of path examples whose relevance to the current query is greater than a specific threshold is selected and retained as query-relevant paths, denoted as . And use it as a supervisory signal to guide the multi-path inferencer to generate multifaceted explanations; Step 3: Establish positive and negative path signals, and use the query-related path examples found in Step 2. As a training triplet The strong positive signal is then used to replace the last product entity in each path with a negative example, thus obtaining the set of negative example paths. , as a strong negative signal of the triple.
5. The product search method for querying relevant multi-faceted interpretability as described in claim 4, characterized in that: The second process specifically involves: using path examples as monitoring signals to infer multipaths, for query-specific triples. The random sampling path serves as a multifaceted explanation for the products retrieved under the current query; that is, a long short-term memory network is first used. The paths are encoded, these path representations are then integrated, and their consistency with the user's current search intent is measured. The probability of matching the user's current search intent is calculated using multiple interpretations, as follows: During the training phase, negative samples with multiple interpretations were constructed for each triple. Multiple explanations of positive samples, including products that the user has never purchased under the current query. The endpoint entities of the path are replaced to obtain new negative sample paths. During training, this part of the problem is treated as a binary classification problem, and the cross-entropy loss is used to calculate the loss in the inference engine. The inference loss is: 。 6. The product search method for querying relevant multi-faceted interpretability as described in claim 5, characterized in that: The training process specifically involves: jointly optimizing the sorter and the multi-path inference engine, with a total loss of: ,in It is a hyperparameter that controls the importance of two components during training; During the inference phase, the following steps are used for each triplet. Generate multifaceted explanations: First, using user... As the starting node, the paths to the goods are sampled, and their meta-paths belong to a predefined set of candidate multipaths. Then, through... Encode each candidate path and calculate its matching probability with the triple: Finally, these candidate paths are sorted in descending order of matching scores, and the paths with scores higher than a set threshold are selected as the final multivariate interpretation.