Personalized product search method and system fusing attribute preference and group feedback

Abstract

The invention discloses a personalized product search method and system fusing attribute preference and group feedback, and belongs to the technical field of deep learning. Existing personalized search algorithms are often interactively driven and neglect scoring information of users, multi-attribute scoring scenes are insufficient in research, fine-grained attribute preferences of the users cannot be effectively modeled, and accordingly user requirements are not accurately grasped. In addition, most algorithms are single in sorting method for the search results and lack necessary feedback information, so that the trust degree of the user for the search results is not favorably improved. In order to solve the problems, a personalized product search method fusing attribute preference and group feedback is provided, a user-query-product weighted hypergraph is constructed based on multi-dimensional scoring information of users, node characterization is learned through hypergraph convolution, and meanwhile, the attribute preference of the users is mined by establishing a level consistency measure. And the search list is rearranged in combination with the multi-attribute score of the group, so that a high-quality search list with higher reference value is provided for the user, and the user satisfaction is improved.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and specifically relates to a personalized product search method and system that integrates attribute preference and group feedback. Background Art

[0002] Personalized product search aims to return a list of relevant products to users based on their query intent and individual preferences. Most early personalized product search algorithms used text embedding technology to convert users, queries, products, and related comments into vector representations, and then obtained prediction scores through inner product operations to generate a ranked list of products. However, this type of algorithm ignores the complex structural relationships between different types of nodes, resulting in poor search results. To this end, people model the interactive relationship between users, queries, and products as a graph structure, and use graph neural networks to learn user preferences and product features to enhance the quality of node representation. In order to further capture the high-order interactive information between the three types of nodes, users, queries, and products, while ensuring the integrity of the relationship between the three, some scholars use hypergraph networks to learn the global representation of various types of nodes to enhance search results.

[0003] However, most existing personalized product search algorithms mine user preferences from the list of products purchased by users, without considering the role of rating information, and rarely involve multi-attribute rating scenarios. They are unable to model user attribute preferences and fine-grained representations, resulting in an inaccurate grasp of user needs. In addition, most existing algorithms simply sort products based on predicted scores, lacking necessary feedback information, which is not conducive to improving users' trust in search results. Summary of the invention

[0004] The purpose of the present invention is to solve the above problems and propose a personalized product search method and system that integrates attribute preferences and group feedback. The total score and the score of each attribute are respectively injected into the user-query-product triplet, and a weighted hypergraph in multiple dimensions is constructed. The representation of nodes in different dimensions is learned by hypergraph convolution, and user preferences and product features are modeled comprehensively and accurately. The influence of each attribute on the total score is calculated using the level consistency measure, so as to achieve an accurate characterization of user attribute preferences. Then, the initial product list is rearranged in combination with group feedback to generate high-quality search results.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A personalized product search method integrating attribute preference and group feedback includes the following steps:

[0007] Step 1: Generate initial representations of users, queries, and products through the Embedding method;

[0008] Step 2: Inject the total score and each attribute score into the user-query-product triple respectively, construct a weighted hypergraph in multiple dimensions, and then use hypergraph convolution to learn the representation of users and products in different dimensions;

[0009] Step 3: Divide the user's rating information into three levels: high, medium, and low, and use the level consistency measure to model the user's attribute preference;

[0010] Step 4: Fuse the representations of the user and product under each attribute respectively, and concatenate them with the representation under the total score to obtain the comprehensive representation of the user and product;

[0011] Step 5: Linearly combine the query representation and the user comprehensive representation, calculate the inner product between the query representation and the product comprehensive representation to obtain the product prediction score, form the initial product list, and then use the cross entropy loss function to optimize the method;

[0012] Step 6: Use the user's attribute preferences to weight the group's multi-attribute ratings of the product into a group comprehensive score, and multiply it with the predicted score to obtain the product's comprehensive recommendation degree, and re-sort the product list accordingly.

[0013] Furthermore, the implementation method of step 1 is as follows:

[0014] Use the Embedding method to encode the user ID, query words, product ID and related comments respectively to obtain the initial representation of the user Query Representation q , and initial characterization of the product

[0015] Furthermore, the implementation method of step 2 is as follows:

[0016] The representation learning of users and products includes two parts: representation learning based on total scores and representation learning based on scores of each attribute;

[0017] In the representation learning part based on the total score, the user's historical interaction records and total score information are combined to construct a weighted hypergraph G containing user-query-product t =(V,H t ), where V = U ∪ Q ∪ I is the node set, H t It is a set of hyperedges. Each hyperedge corresponds to a historical interaction record of the user. It contains three types of nodes: user, query, and product. The weight of the hyperedge is the total score of the user to the product. Then, the hypergraph convolution is used to aggregate the information of nodes and hyperedges to learn the representation of users and products under the total score. The specific process is as follows:

[0018] First, in the lIn the layer hypergraph network, the feature interaction between nodes on the hyperedge is modeled, and the calculation formula is as follows:

[0019]

[0020] Among them, o1, o2, and o3 represent the 1st, 2nd, and 3rd order feature interactions between nodes, respectively, || represents the concatenation operation, and ⊙ represents the element-by-element multiplication operation between vectors;

[0021] Then, the features of each order are interactively spliced ​​to obtain the hyperedge representation The calculation formula is as follows:

[0022]

[0023] Among them, W1 is a trainable parameter matrix;

[0024] Then, the node representation is updated through hyperedge aggregation. For user u, the hyperedge h connected to it is t The normalized weights are:

[0025]

[0026] in, represents the set of all hyperedges connected to user u, Represents the hyperedge h t The corresponding total score;

[0027] Using the normalized hyperedge weight All hyperedges connected to user y are weighted fused to obtain l User representation after layer hypergraph convolution update The calculation formula is as follows:

[0028]

[0029] Finally, the user representations of each layer are concatenated to obtain the representation e of user u under the total score u :

[0030]

[0031] Among them, W2 is a trainable parameter matrix;

[0032] Similarly, the representation e of product i under the total score i The calculation formula is as follows:

[0033]

[0034] Among them, W3 is a trainable parameter matrix;

[0035] In the attribute scoring-based representation learning part, each attribute scoring information is injected into the user-query-product triple to construct a weighted hypergraph under each attribute. Where V = U ∪ Q ∪ I is the node set, is a set of hyperedges, each of which corresponds to a historical interaction record of a user, including three types of nodes: user, query, and product. The weight of the hyperedge is the user's interaction with the product on attribute a. k Ratings below;

[0036] Then, we use formulas (1) to (4) to aggregate node information to obtain hyperedge representation, and replace the total score in formula (5) with the attribute score to calculate the normalized hyperedge weight. Finally, we use formula (6) to aggregate hyperedge information to update node representation, and concatenate the output results of each layer obtained by hypergraph convolution to obtain the user u and product i in attribute a. k The following representation The specific formula is as follows:

[0037]

[0038] Among them, W4 and W5 are trainable parameter matrices.

[0039] Furthermore, the implementation method of step 3 is as follows:

[0040] The user's rating information is divided into three levels: high, medium, and low, corresponding to 3, 2, and 1 respectively. The user's attribute preference is modeled by calculating the consistency measure of each attribute score and the level under the total score;

[0041] In an interaction record, let user u’s rating of product i under the total rating be In attribute a k The corresponding rating levels are Then in this interaction record, attribute a k On the grading consistency measure of the total score The definition is as follows:

[0042]

[0043] For user u, attribute a in all interaction records k The average value of the level consistency measure is obtained to obtain the attribute a of user u k Overall agreement on the total score

[0044]

[0045] Among them, I u represents the set of products that user u has interacted with;

[0046] The higher the overall consistency, the greater the attribute weight. k The weight calculation formula for user u is as follows:

[0047]

[0048] The user's attribute preference is the attribute weight vector w of user u u :

[0049]

[0050] Furthermore, the implementation method of step 4 is as follows:

[0051] For user u, first use its attribute preference to perform weighted fusion on the representations of each attribute to obtain the fused user representation

[0052]

[0053] For product i, average pooling is used to perform fusion to obtain the fused product representation

[0054]

[0055] Then, the representations of users and products under the total score are concatenated with the representations after multi-attribute fusion to obtain the comprehensive representations of users and products.

[0056]

[0057] Among them, W6 and W7 are trainable parameter matrices.

[0058] Furthermore, the implementation method of generating the prediction score in step 5 is as follows:

[0059] First, the user comprehensive representation and query representation are combined q Perform linear combination to get M uq :

[0060]

[0061] Among them, λ is a hyperparameter used to control the ratio between the user comprehensive representation and the query representation;

[0062] Then, for M uq Perform inner product operation with the comprehensive representation of the product to obtain the prediction score according to Sort the products to get the initial product sorting list, The calculation formula is as follows:

[0063]

[0064] After obtaining the predicted score, the cross entropy loss function is used to optimize the method. The specific formula is as follows:

[0065]

[0066] Among them, y uqi Indicates whether there is an interaction between u and i. If there is an interaction, y uqi is 1 if the value is true, otherwise it is 0.

[0067] Furthermore, the specific optimization method in step 6 is as follows:

[0068] For the user group U that has interacted with product i in the initial product list i , average the multi-attribute scores of all users to get the group multi-attribute score Combined with the attribute preference of the current user u, the group multi-attribute score is weighted and integrated to obtain the group comprehensive score of the product.

[0069]

[0070] Among them, u′∈U i are users who have interacted with product i in the list, Represents user u′’s preference for product i on attribute a k The following ratings, Indicates attribute a of user u k The weight of

[0071] The comprehensive recommendation It is defined as the product of the predicted score of the product and the comprehensive score of the group. The calculation formula is as follows:

[0072]

[0073] according to The initial product list is rearranged to obtain the final product sorted list.

[0074] A personalized product search system that integrates attribute preferences and group feedback, including the following modules:

[0075] Initial representation embedding module: Use the Embedding method to encode the user ID, query words, product ID and related comments respectively to obtain the user's initial representation Query Representation q , and initial characterization of the product

[0076] Node representation learning module: inject the total score and each attribute score into the user-query-product triple respectively to construct a weighted hypergraph G in multiple dimensions t and Then, hypergraph convolution is used to learn the representation of users and products under the total rating. u and e i , and both in attribute a k The following representation and

[0077] Attribute preference learning module: divides the user's rating information into three levels: high, medium, and low, and calculates the total rating. Ratings for each attribute The consistency measure between Modeling user attribute preferences w u ;

[0078] Multidimensional representation fusion module: using user attribute preferences w u Characterize the user under each attribute Weighted fusion is At the same time, the average pooling strategy is used to represent the product under each attribute Fusion Then, the representation of the user and product under the total score is spliced ​​with the representation after multi-attribute fusion to obtain the comprehensive representation of the user and product.

[0079] Initial product ranking module: Represent the query as q Comprehensive characterization with users Perform linear combination to get M uq , by calculating its comprehensive characterization with the product The inner product of the product is the predicted score of the product This forms an initial product list, and then the method is optimized using the cross entropy loss function;

[0080] Re-ranking module combining group feedback: The user group that has interacted with the products in the initial product ranking list is recorded as U i , for U i The multi-attribute scores of all users in the group are averaged to obtain the multi-attribute score of the group Then use the user's attribute preference w u Scoring the multi-attribute scores of groups Weighted fusion for group comprehensive score and compare it with the predicted score Multiply to get the comprehensive recommendation of the product Based on this, the product list is reordered to generate a final product sorting list.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] 1. The present invention expands the application scenarios of personalized product search, can make full use of users' multi-dimensional rating information, more comprehensively and accurately capture users' personalized needs and preferences, provide users with more valuable high-quality search lists, and improve users' experience and satisfaction;

[0083] 2. The present invention is applicable to a variety of network applications such as e-commerce platforms, travel booking platforms, health and medical services, etc. It can assist users in making efficient decisions and reduce selection costs by improving the accuracy of search results and the service quality of the platform, thereby improving user stickiness and platform influence. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a schematic diagram of the process of the method recommended by the present invention;

[0085] Figure 2 It is a schematic diagram of the framework of the method recommended by the present invention;

[0086] Figure 3 Schematic diagram of the structure of the system recommended by the present invention. DETAILED DESCRIPTION

[0087] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0088] like Figure 1-3 As shown, the personalized product search method integrating attribute preference and group feedback described in this embodiment includes the following steps:

[0089] Step 1: Generate the initial representation of users, queries, and products through the Embedding method. The implementation method is as follows:

[0090] The present invention uses the embedding method to encode the user ID, the words of the query term, the product ID and its related comments respectively to obtain the initial representation of the user Query Representation q , and initial characterization of the product

[0091] Step 2: Inject the total score and each attribute score into the user-query-product triple respectively, build a weighted hypergraph in multiple dimensions, and then use hypergraph convolution to learn the representation of users and products in different dimensions. The implementation method is as follows:

[0092] The representation learning of users and products includes two parts: the representation learning integrating the total score and the representation learning integrating the scores of each attribute;

[0093] In the representation learning part of the fusion total score, the user's historical interaction records and total score information are combined to construct a weighted hypergraph G containing user-query-product t =(V,H t ), where V = U ∪ Q ∪ I is the node set, H t It is a set of hyperedges. Each hyperedge corresponds to a historical interaction record of the user. It contains three types of nodes: user, query, and product. The weight of the hyperedge is the total score of the user to the product. Then, the hypergraph convolution is used to aggregate the information of nodes and hyperedges to learn the representation of users and products under the total score. The specific process is as follows:

[0094] First, the feature interaction between nodes on the hyperedge is modeled in the l-th layer hypergraph network. The calculation formula is as follows:

[0095]

[0096] Among them, o1, o2, and o3 represent the 1st, 2nd, and 3rd order feature interactions between nodes, respectively, || represents the concatenation operation, and ⊙ represents the element-by-element multiplication operation between vectors;

[0097] Then, the features of each order are interactively spliced ​​to obtain the hyperedge representation The calculation formula is as follows:

[0098]

[0099] Among them, W1 is a trainable parameter matrix;

[0100] Then, the node representation is updated through hyperedge aggregation. For user u, the hyperedge h connected to it is t The normalized weights are:

[0101]

[0102] in, represents the set of all hyperedges connected to user u, Represents the hyperedge h t The corresponding total score;

[0103] Using the normalized hyperedge weight All hyperedges connected to user u are weighted fused to obtain l User representation after layer hypergraph convolution update The calculation formula is as follows:

[0104]

[0105] Finally, the user representations of each layer are concatenated to obtain the representation e of user u under the total score u :

[0106]

[0107] Among them, W2 is a trainable parameter matrix;

[0108] Similarly, the representation e of product i under the total score i The calculation formula is as follows:

[0109]

[0110] Among them, W3 is a trainable parameter matrix;

[0111] In the representation learning part of fusion attribute scoring, each attribute scoring information is injected into the user-query-product triple to construct a weighted hypergraph under each attribute. Where V = U ∪ Q ∪ I is the node set, is a set of hyperedges, each of which corresponds to a historical interaction record of a user, including three types of nodes: user, query, and product. The weight of the hyperedge is the user's interaction with the product on attribute a. k Ratings below;

[0112] Similar to the representation learning under the total score, hypergraph convolution is used to aggregate node and hyperedge information to learn the representation of users and products under various attributes. First, the node information is aggregated using formulas (1) to (4) to obtain the hyperedge representation. Then, the total score in formula (5) is replaced by the attribute score to calculate the normalized hyperedge weight. Finally, formula (6) is used to aggregate the hyperedge information to update the node representation. The output results of each layer obtained by hypergraph convolution are spliced ​​to obtain the user u and product i under attribute a. k The following representation The specific formula is as follows:

[0113]

[0114] Among them, W4 and W5 are trainable parameter matrices.

[0115] Step 3: Divide the user's rating information into three levels: high, medium, and low, and use the level consistency measurement to model the user's attribute preferences. The implementation method is as follows:

[0116] The user's rating information is divided into three levels: high, medium, and low, corresponding to 3, 2, and 1 respectively. The user's attribute preference is modeled by calculating the consistency measure of each attribute score and the level under the total score;

[0117] In an interaction record, let user u’s rating of product i under the total rating be In attribute a k The corresponding rating levels are Then in this interaction record, attribute a kOn the grading consistency measure of the total score The definition is as follows:

[0118]

[0119] For user u, attribute a in all interaction records k The average value of the level consistency measure is obtained to obtain the attribute a of user u k Overall agreement on the total score

[0120]

[0121] Among them, I u represents the set of products that user u has interacted with;

[0122] The higher the overall consistency, the greater the attribute weight. k The weight calculation formula for user u is as follows:

[0123]

[0124] The user's attribute preference is the attribute weight vector w of user u u :

[0125]

[0126] Step 4: Fuse the representations of the user and product under each attribute respectively, and concatenate them with the representation under the total score to obtain the comprehensive representation of the user and product. The implementation method is as follows:

[0127] For user u, first use its attribute preference to perform weighted fusion on the representations of each attribute to obtain the fused user representation

[0128]

[0129] For product i, average pooling is used to perform fusion to obtain the fused product representation

[0130]

[0131] Then, the representations of users and products under the total score are concatenated with the representations after multi-attribute fusion to obtain the comprehensive representations of users and products.

[0132]

[0133] Among them, W6 and W7 are trainable parameter matrices.

[0134] Step 5: Linearly combine the query representation and the user comprehensive representation, calculate the inner product between the query representation and the product comprehensive representation to obtain the product prediction score, form an initial product list, and then use the cross entropy loss function to optimize the method. The implementation method is as follows:

[0135] First, the user comprehensive representation and query representation are combined q Perform linear combination to get M uq :

[0136]

[0137] Among them, λ is a hyperparameter used to control the ratio between the user comprehensive representation and the query representation;

[0138] Then, for M uq Perform inner product operation with the comprehensive representation of the product to obtain the prediction score according to Sort the products to get the initial product sorting list, The calculation formula is as follows:

[0139]

[0140] After obtaining the predicted score, the cross entropy loss function is used to optimize the method. The specific formula is as follows:

[0141]

[0142] Among them, y uqi Indicates whether there is an interaction between u and i. If there is an interaction, y uqi is 1 if the value is true, otherwise it is 0.

[0143] Step 6: Use the user's attribute preferences to weight the group's multi-attribute ratings of the product into a group comprehensive score, and multiply it with the predicted score to get the product's comprehensive recommendation degree. Reorder the product list based on this. The implementation method is as follows:

[0144] For the user group U that has interacted with product i in the initial product list i , average the multi-attribute scores of all users to get the group multi-attribute score Combined with the attribute preference of the current user u, the group multi-attribute score is weighted and integrated to obtain the group comprehensive score of the product.

[0145]

[0146] Among them, u′∈U i are users who have interacted with product i in the list, Represents user u′’s preference for product i on attribute ak The following ratings, Indicates attribute a of user u k The weight of

[0147] The comprehensive recommendation It is defined as the product of the predicted score of the product and the comprehensive score of the group. The calculation formula is as follows:

[0148]

[0149] according to The initial product list is rearranged to obtain the final product sorted list.

[0150] The present invention provides a personalized product search system integrating attribute preference and group feedback, comprising the following modules:

[0151] Initial representation embedding module: Use the Embedding method to encode the user ID, query words, product ID and related comments respectively to obtain the user's initial representation Query Representation q , and initial characterization of the product

[0152] Node representation learning module: inject the total score and each attribute score into the user-query-product triple respectively to construct a weighted hypergraph G in multiple dimensions t and Then, hypergraph convolution is used to learn the representation of users and products under the total rating. u and e i , and both in attribute a k The following representation and

[0153] Attribute preference learning module: divides the user's rating information into three levels: high, medium, and low, and calculates the total rating. Ratings for each attribute The consistency measure between Modeling user attribute preferences w u ;

[0154] Multidimensional representation fusion module: using user attribute preferences w u Characterize the user under each attribute Weighted fusion is At the same time, the average pooling strategy is used to represent the product under each attribute Fusion Then, the representation of the user and product under the total score is spliced ​​with the representation after multi-attribute fusion to obtain the comprehensive representation of the user and product.

[0155] Initial product ranking module: Represent the query as q Comprehensive characterization with users Perform linear combination to get M uq , by calculating its comprehensive characterization with the product The inner product of the product is the predicted score of the product This forms an initial product list, and then the method is optimized using the cross entropy loss function;

[0156] Re-ranking module combining group feedback: The user group that has interacted with the products in the initial product ranking list is recorded as U i , for U i The multi-attribute scores of all users in the group are averaged to obtain the multi-attribute score of the group Then use the user's attribute preference w u Scoring the multi-attribute scores of groups Weighted fusion for group comprehensive score and compare it with the predicted score Multiply to get the comprehensive recommendation of the product Based on this, the product list is reordered to generate a final product sorting list.

Claims

1. A personalized product search method integrating attribute preference and group feedback, characterized in that: The following steps are involved: Step 1: Generate initial representations of users, queries, and products through the Embedding method; Step 2: Inject the total score and each attribute score into the user-query-product triple respectively, construct a weighted hypergraph in multiple dimensions, and then use hypergraph convolution to learn the representation of users and products in different dimensions; Step 3: Divide the user's rating information into three levels: high, medium, and low, and use the level consistency measure to model the user's attribute preference; Step 4: Fuse the representations of the user and product under each attribute respectively, and concatenate them with the representation under the total score to obtain the comprehensive representation of the user and product; Step 5: Linearly combine the query representation and the user comprehensive representation, calculate the inner product between the query representation and the product comprehensive representation to obtain the product prediction score, form the initial product list, and then use the cross entropy loss function to optimize the method; Step 6: Use the user's attribute preferences to weight the group's multi-attribute ratings of the product into a group comprehensive score, and multiply it with the predicted score to obtain the product's comprehensive recommendation degree, and re-sort the product list accordingly.

2. The personalized product search method integrating attribute preference and group feedback according to claim 1, characterized in that: The implementation method of step 1 is as follows: Use the Embedding method to encode the user ID, query words, product ID and its comments respectively to obtain the user's initial representation Query Representation q , and initial characterization of the product 3. The personalized product search method integrating attribute preference and group feedback according to claim 1, characterized in that: The implementation method of step 2 is as follows: The representation learning of users and products includes two parts: representation learning based on total scores and representation learning based on scores of each attribute; In the representation learning part based on the total score, the user's historical interaction records and total score information are combined to construct a weighted hypergraph G containing user-query-product t =(V,H t ), where V = U ∪ Q ∪ I is the node set, H t It is a set of hyperedges. Each hyperedge corresponds to a historical interaction record of the user. It contains three types of nodes: user, query, and product. The weight of the hyperedge is the total score of the user to the product. Then, the hypergraph convolution is used to aggregate the information of nodes and hyperedges to learn the representation of users and products under the total score. The specific process is as follows: First, the feature interaction between nodes on the hyperedge is modeled in the l-th layer hypergraph network. The calculation formula is as follows: Among them, o1, o2, and o3 represent the 1st, 2nd, and 3rd order feature interactions between nodes, respectively, || represents the concatenation operation, and ⊙ represents the element-by-element multiplication operation between vectors; Then, the features of each order are interactively spliced ​​to obtain the hyperedge representation The calculation formula is as follows: Among them, W1 is a trainable parameter matrix; Then, the node representation is updated through hyperedge aggregation. For user u, the hyperedge h connected to it is t The normalized weights are: in, represents the set of all hyperedges connected to user u, Represents the hyperedge h t The corresponding total score; Using the normalized hyperedge weight All hyperedges connected to user u are weighted fused to obtain the updated user representation after the l-th layer hypergraph convolution The calculation formula is as follows: Finally, the user representations of each layer are concatenated to obtain the representation e of user u under the total score u : Among them, W2 is a trainable parameter matrix; Similarly, the representation e of product i under the total score i The calculation formula is as follows: Among them, W3 is a trainable parameter matrix; In the attribute scoring-based representation learning part, each attribute scoring information is injected into the user-query-product triple to construct a weighted hypergraph under each attribute. Where V = U ∪ Q ∪ I is the node set, is a set of hyperedges, each of which corresponds to a historical interaction record of a user, including three types of nodes: user, query, and product. The weight of the hyperedge is the user's interaction with the product on attribute a. k Ratings below; Then, we use formulas (1) to (4) to aggregate node information to obtain hyperedge representation, and replace the total score in formula (5) with the attribute score to calculate the normalized hyperedge weight. Finally, we use formula (6) to aggregate hyperedge information to update node representation, and concatenate the output results of each layer obtained by hypergraph convolution to obtain the user u and product i in attribute a. k The following representation The specific formula is as follows: Among them, W4 and W5 are trainable parameter matrices.

4. The personalized product search method integrating attribute preference and group feedback according to claim 1, characterized in that: The implementation method of step 3 is as follows: The user's rating information is divided into three levels: high, medium, and low, corresponding to 3, 2, and 1 respectively. The user's attribute preference is modeled by calculating the consistency measure of each attribute score and the level under the total score; In an interaction record, let user u’s rating of product i under the total rating be In attribute a k The corresponding rating levels are Then in this interaction record, attribute a k On the grading consistency measure of the total score The definition is as follows: For user u, attribute a in all interaction records k The average value of the level consistency measure is obtained to obtain the attribute a of user u k Overall agreement on the total score Among them, I u represents the set of products that user u has interacted with; The higher the overall consistency, the greater the attribute weight. k The weight calculation formula for user u is as follows: The user's attribute preference is the attribute weight vector w of user u u :

5. The personalized product search method integrating attribute preference and group feedback according to claim 1, characterized in that: The implementation method of step 4 is as follows: For user u, first use its attribute preference to perform weighted fusion on the representations of each attribute to obtain the fused user representation For product i, average pooling is used to perform fusion to obtain the fused product representation Then, the representations of users and products under the total score are concatenated with the representations after multi-attribute fusion to obtain the comprehensive representations of users and products. Among them, W6 and W7 are trainable parameter matrices.

6. The personalized product search method integrating attribute preference and group feedback according to claim 1, characterized in that: The implementation method of step 5 is as follows: First, the user comprehensive representation and query representation are combined q Perform linear combination to get M uq : Among them, λ is a hyperparameter used to control the ratio between the user comprehensive representation and the query representation; Then, for M uq Perform inner product operation with the comprehensive representation of the product to obtain the prediction score according to Sort the products to get the initial product sorting list, The calculation formula is as follows: After obtaining the predicted score, the cross entropy loss function is used to optimize the method. The specific formula is as follows: Among them, y uqi Indicates whether there is an interaction between u and i. If there is an interaction, y uqi is 1 if the value is true, otherwise it is 0.

7. The personalized product search method integrating attribute preference and group feedback according to claim 1, characterized in that: The implementation method of step 6 is as follows: For the user group U that has interacted with product i in the initial product list i , average the multi-attribute scores of all users to get the group multi-attribute score Combined with the attribute preference of the current user u, the group multi-attribute score is weighted and integrated to obtain the group comprehensive score of the product. Among them, u′∈U i are users who have interacted with product i in the list, Represents user u′’s preference for product i on attribute a k The following ratings, Indicates attribute a of user u k The weight of The comprehensive recommendation It is defined as the product of the predicted score of the product and the comprehensive score of the group. The calculation formula is as follows: according to The initial product list is rearranged to obtain the final product sorted list.

8. A personalized product search system integrating attribute preference and group feedback, characterized in that: Includes the following modules: Initial representation embedding module: Use the Embedding method to encode the user ID, query words, product ID and related comments respectively to obtain the user's initial representation Query Representation q , and initial characterization of the product Node representation learning module: inject the total score and each attribute score into the user-query-product triple respectively to construct a weighted hypergraph G in multiple dimensions t and Then, hypergraph convolution is used to learn the representation of users and products under the total rating. u and e i , and both in attribute a k The following representation and Attribute preference learning module: divides the user's rating information into three levels: high, medium, and low, and calculates the total rating. Ratings for each attribute The consistency measure between Modeling user attribute preferences w u ; Multidimensional representation fusion module: using user attribute preferences w u Characterize the user under each attribute Weighted fusion is At the same time, the average pooling strategy is used to represent the product under each attribute Fusion Then, the representation of the user and product under the total score is spliced ​​with the representation after multi-attribute fusion to obtain the comprehensive representation of the user and product. Initial product ranking module: Represent the query as q Comprehensive characterization with users Perform linear combination to get M uq , by calculating its comprehensive characterization with the product The inner product of the product is the predicted score of the product This forms an initial product list, and then the method is optimized using the cross entropy loss function; Re-ranking module combining group feedback: The user group that has interacted with the products in the initial product ranking list is recorded as U i , for U i The multi-attribute scores of all users in the group are averaged to obtain the multi-attribute score of the group Then use the user's attribute preference w u Scoring the multi-attribute scores of groups Weighted fusion for group comprehensive score and compare it with the predicted score Multiply to get the comprehensive recommendation of the product Based on this, the product list is reordered to generate a final product sorting list.

Images