Personalized search method and system for group enhancement

Abstract

The invention discloses a group-enhanced personalized search method and system, and relates to the field of document search. The method specifically comprises the following steps: obtaining a short-term static interest vector of a user according to a plurality of historical queries of a current query and a click document list; obtaining a long-term static interest vector of the user according to the historical session; obtaining a short-term dynamic interest vector and a long-term dynamic interest vector of the user based on the query; based on the current query of the target user and the user representation vector, finding a plurality of similar related user groups to be matched with the candidate document list; performing correlation scoring on the number of clicks of the candidate document list in the current query, the number of clicks of the candidate document list in the search history, the click entropy of the candidate document list in the current query and the user matching degree of the plurality of related user groups and the candidate document list; the list of candidate documents is ranked based on the relevance score. According to the method, each query can be distinguished by combining the user query and the document feature representation, and the performance of personalized dynamic search is improved by fusing group user information.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of document search, in particular to a group-enhanced personalized search method and system. BACKGROUND

[0002] For the same query request, many search engines will return the same document list. However, for different users, the same query request may represent completely different information needs, and personalized search is thus proposed and widely applied. Personalized search provides a more personalized ranking result when returning search results according to the user's interests, thereby improving the quality of search results and user satisfaction. Many existing personalized search models mainly rely on extracting features such as clicks, topics and topology from user search history to analyze user interests. Specifically, when a user initiates a query, the system will use the user's past search history log to assist in realizing personalized ranking of search results. The user's search history log contains a series of query records and clicked documents. Obviously, not all historical interactions contribute equally to the ranking of candidate documents, so some schemes consider introducing a group of users with common interests to enhance personalized search. Existing group-enhanced personalized search schemes usually enrich user portraits by introducing a group of users with common interests. These methods mainly use artificially designed rules (such as commonly clicked documents) to determine the user group. Although these grouping methods have achieved certain improvement effect, there are still some limitations. The existing methods have the following shortcomings: first, they use artificially designed features (such as commonly clicked documents) to find a group of users with common interests to enhance the performance of personalized search, but these artificially designed features have poor scalability and are time-consuming and labor-intensive. Second, the existing methods can only obtain the user interests covered by limited features, ignoring the discriminative information between each query in the search history. SUMMARY

[0003] The technical problem to be solved by the present application is that artificially designed features have poor scalability and are time-consuming and labor-intensive, and can only obtain limited user interests, ignoring the discriminative information between each query in the search history. The purpose is to provide a group-enhanced personalized search method and system, which is different from the traditional static features that rely on click behavior and the query-independent user grouping method. The present application combines user queries and document feature representations to distinguish the deep semantic information of each query, thereby fusing group user information to improve the performance of personalized dynamic search, solving the above technical problems.

[0004] The present application is implemented by the following technical solutions:

[0005] A personalized search method with group enhancement, comprising: obtaining search histories of multiple users; the search histories comprising multiple historical sessions; each of the historical sessions comprising multiple queries and a clicked document list corresponding to each of the queries; learning a user representation vector from the clicked document list, specifically comprising: obtaining a short-term static interest vector of a user according to multiple historical queries and the clicked document list of a current session when the user issues the query for the first time; obtaining a long-term static interest vector of the user according to the historical sessions of the user based on the perception of the short-term static interest vector; obtaining a short-term dynamic interest vector of the user according to the current session of the user based on the perception of the query; obtaining a long-term dynamic interest vector of the user according to the historical sessions of the user based on the perception of the query; returning a candidate document list for a current query initiated by a target user by using an existing search engine; finding multiple relevant user groups based on the current query of the target user and the user representation vector; weighting and fusing the user representation vectors of the multiple relevant user groups; matching the fused user representation vector with the candidate document list; performing relevance scoring on the candidate document list according to the clicked document list corresponding to the current query of the target user, the search histories, and the user representation vectors of the multiple relevant user groups, specifically comprising: inputting the current number of clicks of the candidate document list in the current query of the target user, the historical number of clicks of the candidate document list in the search histories, the click entropy of the candidate document list in the current query, and the user matching degree of the multiple relevant user groups and the candidate document list into a multi-layer perception machine to perform relevance scoring on the candidate document list of the target user; and reordering the candidate document list based on the relevance scoring.

[0006] The short-term static interest vector of the user when the user issues the query for the first time according to the multiple historical queries and the clicked document list of the current session, specifically comprising: establishing a hierarchical recurrent neural network model; the hierarchical recurrent neural network model comprising an intra-session neural network level; the input of the intra-session neural network being the multiple queries of a single historical session and the clicked document list corresponding to each of the queries, and the output being a hidden state vector of a session representation; the hidden state vector of each session being calculated according to the current input and the state vector of the previous query; and the hidden state vector of the current query being the short-term interest vector of the user when the user issues the query for the first time.

[0007] The above perception based on the short-term static interest vector obtains a long-term static interest vector of the user according to the historical conversation of the user, and specifically includes that the hierarchical recurrent neural network model includes an inter-conversation neural network; the levels of the intra-conversation neural network and the inter-conversation neural network are arranged from low to high; the input of the inter-conversation neural network is a single historical conversation, and the output is a sequence of short-term interest vectors of the user at the end of each historical conversation; a attention mechanism layer is applied on the hidden state vector of the inter-conversation neural network according to the current query to obtain the sequence of short-term interest vectors of the user as the long-term static interest vector.

[0008] The above perception based on the query obtains a short-term dynamic interest vector of the user according to the current conversation of the user, and includes that the intra-conversation neural network introduces the short-term static interest vector with a perception attention mechanism of the current query, and defines the short-term static interest vector as the short-term dynamic interest vector.

[0009] The above perception based on the query obtains a long-term dynamic interest vector of the user according to the historical conversation of the user, and includes that the inter-conversation neural network introduces the long-term static interest vector with a perception attention mechanism of the current query, and defines the long-term static interest vector as the long-term dynamic interest vector.

[0010] The above perception based on the current query and the user representation vector of the target user finds a plurality of relevant user groups, and specifically includes that based on the current query, the short-term static interest vector, the long-term static interest vector, the short-term dynamic interest vector and the long-term dynamic interest vector, the cosine similarity between the target user and different users is calculated respectively, so that the first relevant user group corresponding to the current query, the second relevant user group corresponding to the short-term static interest vector, the third relevant user group corresponding to the long-term static interest vector, the fourth relevant user group corresponding to the short-term dynamic interest vector and the fifth relevant user group corresponding to the long-term dynamic interest vector are ranked according to the cosine similarity score;

[0011] The above user representation vectors of the plurality of relevant user groups are weighted and fused, and specifically include that five multi-head attention layers are used to weightedly fuse the first relevant user group, the second relevant user group, the third relevant user group, the fourth relevant user group and the fifth relevant user group.

[0012] The method further comprises: collecting a plurality of training sample pairs of user-clicked relevant documents and user-not-clicked irrelevant documents of the same candidate document list; exchanging the ranking of the user-clicked relevant documents and the user-not-clicked irrelevant documents; and establishing a loss function by the cross-entropy between the expected probability of correct prediction of the relevance of the plurality of user-clicked relevant documents and the user-not-clicked irrelevant documents and the prediction probability, and the product of the exchange ranking change.

[0013] A population-enhanced personalized search system comprises: a historical clicked document module configured to obtain search histories of a plurality of users, wherein the search histories comprise a plurality of historical sessions, and each historical session comprises a plurality of queries and a clicked document list corresponding to each query; a clicked document representation module configured to learn user representation vectors from the clicked document list, specifically comprising: obtaining a short-term static interest vector of a user according to a plurality of historical queries and the clicked document list of a current session when the user issues a query for the first time; obtaining a long-term static interest vector of the user according to the historical sessions of the user based on the perception of the short-term static interest vector; obtaining a short-term dynamic interest vector of the user according to the current session of the user based on the perception of the query; obtaining a long-term dynamic interest vector of the user according to the historical sessions of the user based on the perception of the query; a candidate document search module configured to return a candidate document list for a current query initiated by a target user using an existing search engine; a relevant user group module configured to find a plurality of relevant user groups similar to the target user based on the current query of the target user and the user representation vectors; a candidate document matching module configured to: weight and fuse the user representation vectors of the plurality of relevant user groups; and match the fused user representation vectors with the candidate document list; a candidate document scoring module configured to score the relevance of the candidate document list based on the clicked document list corresponding to the current query of the target user, the search histories, and the user representation vectors of the plurality of relevant user groups, specifically comprising: inputting the current number of clicks of the candidate document list in the current query of the target user, the historical number of clicks of the candidate document list in the search histories, the click entropy of the candidate document list in the current query, and the user matching degree of the plurality of relevant user groups and the candidate document list into a multilayer perceptron to score the relevance of the candidate document list of the target user; and a candidate document ranking module configured to reorder the candidate document list based on the relevance score.

[0014] An electronic device includes a memory, a processor, and a computer program running on the processor, wherein the processor executes the computer program to implement the steps of the group-enhanced personalized search method described above.

[0015] A computer-readable storage medium storing a computer program, characterized in that: when the computer program is executed by a processor, it implements the steps of the aforementioned population-enhanced personalized search method.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0017] This invention provides a group-enhanced personalized search method. Specifically, when a user first submits the aforementioned query, a short-term static interest vector is obtained based on multiple historical queries and clicked document lists in the current session. Based on the perception of the short-term static interest vector, a long-term static interest vector is obtained based on the user's historical sessions. Based on the perception of the query, a short-term dynamic interest vector is obtained based on the user's current session. Based on the perception of the query, a long-term dynamic interest vector is obtained based on the user's historical sessions. This allows the learning of short-term and query-perceived dynamic user representation vectors from the clicked document list. Based on the target user's current query and the obtained user representation vectors, multiple relevant user groups similar to the target user are identified. The user representation vectors of multiple relevant user groups are weighted and fused, and then matched with the candidate document list found by the target user's current query to obtain the user group features corresponding to the candidate documents. The current click count of the candidate document list in the target user's current query, the historical click count in the search history, the click entropy of the candidate document list in the current query, and the user matching degree between multiple relevant user groups and the candidate document list are input into a multilayer perceptron to score the relevance of the target user's candidate document list. This invention learns user characteristics based on the historical clicked document lists of users and related user groups, and combines this with the distinctive information of each query to capture deeper connections between user groups, thereby dynamically integrating group user information to improve the performance of personalized search. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0019] Figure 1 This is a flowchart illustrating a group-enhanced personalized search method as described in this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this invention are only for explaining this invention and are not intended to limit this invention.

[0021] Example

[0022] like Figure 1 As shown, this application provides a group-enhanced personalized search method, including: acquiring the search history of multiple users; the search history includes multiple historical sessions; each historical session contains multiple queries and a list of clicked documents corresponding to each query; learning user representation vectors from the list of clicked documents, specifically including: when a user first issues the query, obtaining the user's short-term static interest vector based on the multiple historical queries of the current session and the list of clicked documents; based on the perception of the short-term static interest vector, obtaining the user's long-term static interest vector based on the user's historical sessions; based on the perception of the query, obtaining the user's short-term dynamic interest vector based on the user's current session; based on the perception of the query, obtaining the user's long-term dynamic interest vector based on the user's historical sessions; using an existing search engine to return a list of candidate documents for the current query initiated by the target user; based on the target user's current... The query and the aforementioned user representation vector are used to find multiple similar related user groups; the user representation vectors of the multiple related user groups are weighted and fused; the fused user representation vectors are matched with the aforementioned candidate document list; based on the aforementioned clicked document list corresponding to the current query of the target user, the aforementioned search history, and the user representation vectors of the multiple related user groups, the relevance score of the aforementioned candidate document list is calculated, specifically including: inputting the current click count of the aforementioned candidate document list in the current query of the target user, the historical click count of the aforementioned candidate document list in the aforementioned search history, the click entropy of the aforementioned candidate document list in the current query, and the user matching degree between the multiple related user groups and the aforementioned candidate document list into a multilayer perceptron to calculate the relevance score of the aforementioned candidate document list of the target user; and reordering the aforementioned candidate document list based on the aforementioned relevance score.

[0023] The process involves inputting the current click count of the candidate document list within the clicked document list corresponding to the target user's current query, the historical click count of the candidate document list in the search history, the click entropy of the candidate document list in the current query, and the user matching degree between multiple relevant user groups and the candidate document list into a multilayer perceptron. This data is then weighted and fused to predict the matching weight of each candidate document, and a relevance score is assigned to the target user's candidate document list. The aforementioned user matching degree refers to the percentage of users matching each document, thus enabling branch evaluation of the relevance of each document.

[0024] Specifically, suppose there exists a set A single user, denoted as When a user In the Current time in the next session Initiate a new query At that time, the user has the corresponding search history. This includes both long-term and short-term history. Long-term history consists of previous search sessions. It means that among them It is the index of the current session; the short-term history is the index of the current search session. Therefore, the complete search history can be represented as: Each session is defined as a query sequence and a list of documents corresponding to each query. The candidate document list refers to the search results for documents to be determined using a general search engine for the current query. The clicked document list refers to the search results for clicked documents that the user has confirmed selecting in their search history.

[0025] For the A session, whose general representation is:

[0026] ;

[0027] in, Representing a session The total number of queries. Currently, users Initiated a query The original, non-personalized search engine returned a list of candidate documents. ,in yes Abbreviation yes The abbreviation. (In the user's...) Initiate a query The moment It can be based on some kind of... and Dynamically identify a group of relevant users, denoted as .

[0028] For each document, its relevance score is calculated separately based on the following three aspects, which can be expressed as:

[0029] ;

[0030] Among them, the current query The target user's past search history User groups related to the target users ; Document With query The correlation between them Document Search history of the target user The correlation between them Document With relevant user groups The correlation between them. Based on the overall score. For the candidate document list The documents are then reordered, resulting in a personalized list of sorted documents.

[0031] By designing a group-based personalized search framework, which enhances the personalized search effect for the target user under the current query by utilizing information from a group of related users. Our proposed personalized search model GBPS mainly consists of three parts: (1) a document sequence deep learning model that does not depend on the user group, used to construct the user's long-term and short-term interests; (2) the formation of the user group; and (3) the fusion of user group information to complete the personalized search.

[0032] Specifically, GBPS improved upon the hierarchical recurrent neural network model (HRNN) by creating Extended-HRNN. Extended-HRNN was then used to learn user profiles from clicked documents extracted from the user's search history. HRNN was chosen as the base personalized search model (i.e., the embedded model) because of its superior characteristics. HRNN is a representation-based model designed specifically for personalized search tasks, capable of considering the sequential information inherent in past queries and sessions. It primarily consists of two RNN layers, used to model the sequential information between the initiated query and past sessions, and can automatically generate user profiles from historical data.

[0033] Specifically, HRNN first divides the user's search history into multiple sessions: ,in This represents the index of the current session. Each session contains not only a series of queries but also a list of clicked documents corresponding to each query. Therefore, the index of the current session... The formal representation of a session is:

[0034] ;

[0035] in This indicates the total number of queries performed in this session.

[0036] The representation of a single session is computed using a low-level intra-session RNN. Its input is a series of query-click document pairs, and its output is the latent state vector for that session. The final session representation is taken from the last hidden state vector of this RNN. In the low-level RNN of each session, the hidden state vector at each step It is calculated based on the current input and the state vector of the previous step, as shown in the following formula:

[0037] ;

[0038] When a user submits a query At that time, the last hidden state vector corresponding to the history part of the current session. Considered as a user's short-term interest vector .

[0039] The past session representations from the user's search history are then fed as input to a higher-level inter-session RNN. This higher-level RNN computes the user representation sequence at the end of each session. The formula is as follows:

[0040] ;

[0041] Then, query based on input. An attention mechanism layer is applied to the hidden state vector of a high-level RNN to generate the user's long-term interest vector.

[0042] Extended-HRNN extends the computation of short-term and long-term user interest vectors in HRNN, performing calculations with and without query-aware attention mechanisms, respectively. We introduce a long-term user interest vector with query-aware attention from HRNN and define it as our long-term dynamic interest vector. The formula is expressed as follows:

[0043] ;

[0044] in, This represents an attention layer. As a query in the attention mechanism, It serves as both the key and value in the attention mechanism. Clearly, the last hidden state vector is used directly. This is one of the simplest ways to generate a representation of long-term user interests. Furthermore, this vector is independent of the current query, thus reflecting a relatively objective long-term user interest unaffected by the current query. Based on this, further... As our long-term static interest vector To expand the representation of long-term user interests.

[0045] Similarly, the short-term user interests calculated in the HRNN model are used as our short-term static interest vector. Its value is equal to Furthermore, the query-aware attention mechanism used in high-level RNNs within HRNN models has been shown to dynamically highlight different sessions within a user's historical search behavior. We believe that introducing a query-aware attention mechanism into the low-level RNN of the current session can similarly assign appropriate weights to each query within that session, thereby avoiding unnecessary noise interference. Therefore, we apply a query-aware attention layer to the hidden state vector of the current session to compute our short-term dynamic interest vector. The formula is expressed as follows:

[0046] ;

[0047] in, In this attention layer, it is used as a query, and It can be used as both a key and a value.

[0048] In this way, given a user's search history and current query, four vectors can be constructed to represent the user's interests: a long-term static user profile, a long-term dynamic user profile, a short-term static user profile, and a short-term dynamic user profile. Here, "static" refers to sequential search history that is not modeled using a current query-aware attention mechanism, while "dynamic" corresponds to modeling methods that incorporate a current query-aware attention mechanism.

[0049] For a target user who has just initiated a new query, a user profile is constructed, generating four vectors to represent their interests from four different perspectives: long-term, short-term, dynamic, and static. To reduce the enormous computational and time overhead of each new query, a long-term static interest vector can be used as a static representation of historical users. This interest vector is constructed sequentially and reflects the user's stable preferences. Since no query-aware attention mechanism is used in its construction, it is unbiased and unaffected by any specific query. Therefore, the long-term static interest vector of each user is... Save it to your hard drive and update it regularly.

[0050] The user group segmentation in this application is based on the similarity between the user representation vector and the current query. Conventionally, user groups are constructed based on the similarity between two representation vectors. For a target user... and an input query We constructed four types of user profile representations and simultaneously obtained a vector representing the current query. Furthermore, the initiated query itself can be seen as a strong signal of the user's current interests. Thanks to the structural characteristics of HRNNs, the query representation vector and the four user profile vectors mentioned above reside in the same semantic space. This ensures that it is meaningful to further calculate the similarity between the query representation and the user profile representation, each revealing a different aspect of the target user's interests.

[0051] For another user who needs to be considered It already has a stored static representation, namely, the long-term static interest vector that she most recently calculated and saved. Let... This represents a similarity function used to calculate the similarity between two vectors of the same dimension. and The similarity between them. Experiments have verified that the cosine similarity function outperforms the L2 similarity function (where L2 similarity is calculated as follows). ).

[0052] The GBPS framework will use short-term static interest vectors Long-term static interest vectors Short-term dynamic interest vectors Long-term dynamic interest vectors These four user profile representations and the target user's current query, totaling five representation vectors revealing the target user's interests, are considered as five distinct grouping intentions. Based on these five grouping intentions, the target user... Other users The similarity between them can be calculated in the five forms mentioned above, as follows:

[0053] ;

[0054] in, Indicates time Initiated query The vector representation of the similarity. Each similarity calculation method can group users similar to the target user together to form a subgroup. Therefore, based on the five similarity forms mentioned above, we can identify the target user... And the current query Five user subgroups were identified. Each subgroup was extracted by sorting all other users in descending order of their corresponding similarity scores.

[0055] For a fixed user and in time Recently initiated query It can form five subgroups, each containing For each of the 3 similar users, the formal representation can be as follows:

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] ;

[0061] in, Represents the cosine similarity score The ranking.

[0062] In acquiring target users Initiated query People with shared interests, from five subgroups After finding the nearest neighbor users, load the static user profiles of these users from the hard drive and combine these representation vectors into a matrix.

[0063] This matrix contains information about the target user and all groups retrieved for the current query.

[0064] To fully utilize this group information, five multi-head attention layers can be applied to fuse it. Each multi-head attention layer corresponds to a different grouping intent, and all group information is weighted and fused according to this intent. This results in five centralized representations of the group information. Taking a query as the grouping intent as an example, the centralized representation corresponding to the current query is calculated as follows:

[0065] ;

[0066] In this way, group information is integrated into five centralized representations: These five representations are fed into a multilayer perceptron (MLP) and used... As an activation function.

[0067] In addition to the personalized score calculated by the model, query document relevance features and click behavior features were introduced as additional features. Since the original query document features were unavailable in our dataset, we used the document's position in the original ranking as a substitute feature. Click behavior features included: the total number of historical clicks on candidate documents, the number of clicks on the candidate document in the current query, and the click entropy of the current query. Click entropy refers to the diversity and distribution of user click behavior, which can be used to measure the randomness and disorder of user clicks on search engine results pages. For example, if the distribution of search results clicked by users is relatively uniform, the click entropy is high; if users mainly click on the top-ranked results, the click entropy is low. The reason for introducing click entropy is that the personalization effect varies greatly between different queries; applying personalization indiscriminately to all queries may negatively impact the overall search quality. These additional features were also input into a... In the multilayer perceptron (MLP) as an activation function.

[0068] The framework is trained using a basic ranking algorithm, LambdaRank. Training sample pairs are generated from the query logs: for example, SAT clicked documents are considered relevant documents, and other documents are considered irrelevant. The loss is then calculated using the representation vectors of the document pairs. For a relevant document... and an unrelated document Taking a sample pair as an example, the loss function is defined as: the cross-entropy between the expected probability and the predicted probability when the ranking positions of these two documents are swapped, versus the change in the ranking metric. The product of . The specific formula is as follows:

[0069] ;

[0070] in, Documents representing model predictions Compared to documents More relevant probability, and This represents the true probability. Specifically, the predicted probability is calculated using a logistic function, ensuring it falls between 0 and 1:

[0071] .

[0072] In summary, this application provides a group-enhanced personalized search method and system: The present invention provides a group-enhanced personalized search method, specifically, when a user issues the above query for the first time, it obtains the user's short-term static interest vector based on multiple historical queries and clicked document lists in the current session; based on the perception of the short-term static interest vector, it obtains the user's long-term static interest vector based on the user's historical sessions; based on the perception of the query, it obtains the user's short-term dynamic interest vector based on the user's current session; based on the perception of the query, it obtains the user's long-term dynamic interest vector based on the user's historical sessions; thereby learning the short-term and query-perceived dynamic user representation vectors of the current session from the clicked document list; based on the target user's current query and the obtained user representation vectors, it finds multiple related user groups similar to the target user; after weighted fusion of the user representation vectors of multiple related user groups, it matches them with the candidate document list found in the target user's current query to obtain the user group features corresponding to the candidate documents; and inputs the current click count of the candidate document list in the target user's current query, the historical click count in the search history, the click entropy of the candidate document list in the current query, and the user matching degree between multiple related user groups and the candidate document list into a multilayer perceptron to score the relevance of the target user's candidate document list. This invention learns user characteristics based on the historical clicked document lists of users and related user groups, and combines this with the distinctive information of each query to capture deeper connections between user groups, thereby dynamically integrating group user information to improve the performance of personalized search.

[0073] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A group-enhanced personalized search method, characterized in that, include: Retrieve the search history of multiple users; the search history includes multiple historical sessions; each historical session contains multiple queries and a list of clicked documents corresponding to each query; Learning user representation vectors from the list of clicked documents specifically includes: When a user submits the query for the first time, a short-term static interest vector is obtained based on multiple historical queries in the current session and the list of clicked documents. Based on the perception of the short-term static interest vector, a long-term static interest vector is obtained based on the user's historical sessions. Based on the perception of the query, a short-term dynamic interest vector is obtained based on the user's current session. Based on the perception of the query, a long-term dynamic interest vector is obtained based on the user's historical sessions. The existing search engine is used to return a list of candidate documents for the current query initiated by the target user; Based on the target user's current query and the user representation vector, find multiple similar related user groups; The user representation vectors of multiple related user groups are weighted and fused; the fused user representation vectors are then matched with the candidate document list. The relevance score of the candidate document list is calculated based on the clicked document list corresponding to the target user's current query, the search history, and the user representation vectors of multiple related user groups. Specifically, this includes inputting the current click count of the candidate document list in the target user's current query, the historical click count of the candidate document list in the search history, the click entropy of the candidate document list in the current query, and the user matching degree between multiple related user groups and the candidate document list into a multilayer perceptron to calculate the relevance score of the target user's candidate document list. The candidate document list is reordered based on the relevance score.

2. The group-enhanced personalized search method according to claim 1, characterized in that, When a user submits the query for the first time, a short-term static interest vector is obtained based on multiple historical queries in the current session and the list of clicked documents. This vector specifically includes: A hierarchical recurrent neural network model is established; the hierarchical recurrent neural network model includes intra-session neural network layers; The input to the intra-session neural network is multiple queries of a single historical session and the list of clicked documents corresponding to each query, and the output is a hidden state vector representing the session; the hidden state vector of each session is calculated based on the current input and the state vector of the previous query; when the user issues the query for the first time, the hidden state vector of the current query is the user's short-term interest vector.

3. The group-enhanced personalized search method according to claim 2, characterized in that, The perception based on the short-term static interest vector, and the obtaining of the user's long-term static interest vector based on the user's historical sessions, specifically includes: The hierarchical recurrent neural network model includes an inter-session neural network; the levels of the intra-session neural network and the inter-session neural network are set from low to high. The input to the inter-session neural network is a single historical session, and the output is a sequence of short-term interest vectors of the user at the end of each historical session. An attention mechanism layer is applied to the hidden state vector of the inter-session neural network based on the current query to obtain the user's short-term interest vector sequence as a long-term static interest vector.

4. The group-enhanced personalized search method according to claim 3, characterized in that, The perception based on the query, and the acquisition of the user's short-term dynamic interest vector according to the user's current session, includes: the in-session neural network introducing the short-term static interest vector with the perception attention mechanism of the current query, and defining it as a short-term dynamic interest vector.

5. The group-enhanced personalized search method according to claim 4, characterized in that, The perception based on the query, and the obtaining of the user's long-term dynamic interest vector according to the user's historical sessions, includes: the inter-session neural network introducing the long-term static interest vector with the perception attention mechanism of the current query, and defining it as the long-term dynamic interest vector.

6. The group-enhanced personalized search method according to claim 1, characterized in that, The step of finding multiple similar related user groups based on the target user's current query and the user representation vector specifically includes: calculating the cosine similarity between the target user and different users based on the current query, the short-term static interest vector, the long-term static interest vector, the short-term dynamic interest vector, and the long-term dynamic interest vector, and then ranking them according to the cosine similarity scores to obtain the first related user group corresponding to the current query, the second related user group corresponding to the short-term static interest vector, the third related user group corresponding to the long-term static interest vector, the fourth related user group corresponding to the short-term dynamic interest vector, and the fifth related user group corresponding to the long-term dynamic interest vector; The weighted fusion of the user representation vectors of the multiple related user groups specifically includes: using five multi-head attention layers to perform weighted fusion on the first related user group, the second related user group, the third related user group, the fourth related user group, and the fifth related user group.

7. The group-enhanced personalized search method according to claim 6, characterized in that, The step of inputting the historical total number of clicks on the candidate document list, the number of clicks on the candidate document list under the current query, the click entropy of the current query, and the relevance score of the candidate document list into a multilayer perceptron to predict the relevance score of the candidate document list further includes: collecting multiple training sample pairs of user-clicked relevant documents and user-unclicked irrelevant documents from the same candidate document list; swapping the ranking of the user-clicked relevant documents and the user-unclicked irrelevant documents; and establishing a loss function by multiplying the expected probability and predicted probability of the model correctly predicting the relevance of the multiple sets of user-clicked relevant documents and user-unclicked irrelevant documents with the change in ranking.

8. A group-enhanced personalized search system, characterized in that, include: Historical Clicked Documents Module: Retrieves the search history of multiple users; the search history includes multiple historical sessions; each historical session contains multiple queries and a list of clicked documents corresponding to each query; Click document representation module: Learns user representation vectors from the click document list, specifically including: when the user issues the query for the first time, obtaining the user's short-term static interest vector based on multiple historical queries of the current session and the click document list; based on the perception of the short-term static interest vector, obtaining the user's long-term static interest vector based on the user's historical sessions; based on the perception of the query, obtaining the user's short-term dynamic interest vector based on the user's current session; based on the perception of the query, obtaining the user's long-term dynamic interest vector based on the user's historical sessions. Candidate document search module: Uses existing search engines to return a list of candidate documents for the current query initiated by the target user; Related user group module: Based on the target user's current query and the user representation vector, find multiple similar related user groups; Candidate document matching module: performs weighted fusion of the user representation vectors of multiple related user groups; matches the fused user representation vector with the candidate document list; Candidate document scoring module: Based on the clicked document list corresponding to the target user's current query, the search history, and the user representation vectors of multiple related user groups, the module performs a relevance score on the candidate document list. Specifically, it inputs the current click count of the candidate document list in the target user's current query, the historical click count of the candidate document list in the search history, the click entropy of the candidate document list in the current query, and the user matching degree between multiple related user groups and the candidate document list into a multilayer perceptron to perform a relevance score on the target user's candidate document list. Candidate document sorting module: Reorders the candidate document list based on the relevance score.

9. An electronic device comprising a memory, a processor, and a computer program running on the processor, characterized in that: When the processor executes the computer program, it implements the steps of a group-enhanced personalized search method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of a population-enhanced personalized search method as described in any one of claims 1 to 7.

Images