CN-122019889-A - User multi-interest recommendation method and system based on large model and frequency domain decomposition

CN122019889ACN 122019889 ACN122019889 ACN 122019889ACN-122019889-A

Abstract

The application relates to the field of large model application, in particular to a user multi-interest recommendation method and system based on a large model and frequency domain decomposition. The method comprises the steps of obtaining historical interaction items of a user and objective attribute characteristics of candidate items based on item objective attributes, selecting a plurality of objective interest sequences forming the candidate items from the historical interaction items according to the objective attribute characteristics, processing the candidate items and user comment information of the historical interaction items through a large language model, selecting a plurality of subjective interest sequences forming the candidate items from the historical interaction items, sequentially carrying out time sequence coding, frequency domain decomposition and inverse Fourier transformation on the sequences to obtain time domain representations, respectively fusing the time domain representations corresponding to the sequences to obtain comprehensive interest representations, carrying out preference prediction based on the comprehensive interest representations and the objective attribute characteristics of the candidate items, and outputting item recommendation results of the user. The recommendation method and the recommendation device are used for improving the matching degree of the recommendation result in accuracy and personalized experience.

Inventors

ZHAO JIE
GUO HAOTIAN
ZHANG CUIHONG
WU XIUZHU

Assignees

广东工业大学

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (11)

1. A user multi-interest recommendation method based on a large model and frequency domain decomposition is characterized by comprising the following steps: based on preset objective attributes of the items, objective attribute characteristics of historical interaction items of the user are obtained; Acquiring objective attribute characteristics of candidate items based on the objective attributes of the items; Selecting a plurality of history interaction items from the history interaction items according to the objective attribute characteristics of the candidate items and the objective attribute characteristics of the history interaction items to form an objective interest sequence of the candidate items; Processing the user comment information of the candidate item and the user comment information of the history interaction item through a large language model, and selecting a plurality of items from the history interaction item to form a subjective interest sequence of the candidate item; Sequentially performing time sequence coding, frequency domain decomposition and inverse Fourier transform on the objective interest sequence and the subjective interest sequence to obtain corresponding time domain representation; Respectively fusing time domain representations corresponding to the objective interest sequence and the subjective interest sequence to obtain corresponding comprehensive interest representations; and carrying out preference prediction based on the comprehensive interest expression and the objective attribute characteristics of the candidate items, and outputting item recommendation results of the users.
2. The method according to claim 1, wherein the obtaining objective attribute features of the user's historical interaction items based on preset objective attributes of the items includes: Presetting item objective attribute categories, and presetting a trainable embedded lookup table corresponding to each item objective attribute category; Acquiring the item objective attribute of the history interaction item, and respectively searching corresponding low-dimensional dense vectors from corresponding embedded lookup tables according to the item objective attribute of the history interaction item; Splicing, dimension alignment and dimension mapping are sequentially carried out on the low-dimensional dense vectors of the history interaction items, and a first objective attribute representation vector is obtained; Presetting a plurality of objective interest visual angles, wherein each objective interest visual angle corresponds to a learnable linear projection matrix; Based on the linear projection matrix corresponding to the objective interest visual angle, performing feature mapping on the first objective attribute representation vector to obtain a corresponding initial item representation; The method comprises the steps of sorting initial item representations of the history interaction items under the corresponding objective interest view angles according to interaction time of the corresponding history interaction items and the user to obtain an initial item representation sequence, inputting the initial item representation sequence into a time sequence coding network for processing, and outputting an item representation sequence with enhanced time sequence; And taking the time sequence enhanced item representation sequences of the history interaction items under all objective interest view angles as objective attribute characteristics of the history interaction items, wherein each time sequence enhanced item representation sequence comprises item representations corresponding to each history interaction item under each objective interest view angle.
3. The method of claim 2, wherein the obtaining objective attribute characteristics of candidate items based on the item objective attributes comprises: Acquiring item objective attributes of the candidate items; Searching corresponding low-dimensional dense vectors from corresponding embedded lookup tables according to item objective attributes of the candidate items; splicing, dimension alignment and dimension mapping are sequentially carried out on the low-dimensional dense vectors of the candidate items, and a second objective attribute representation vector is obtained; Based on the linear projection matrix corresponding to the objective interest visual angle, performing feature mapping on the second objective attribute representation vector to obtain a corresponding item representation; and taking the item representation of the candidate item under all objective interest perspectives as the objective attribute characteristic of the candidate item.
4. The method according to claim 3, wherein selecting a plurality of history interaction items from the history interaction items according to the objective attribute characteristics of the candidate items and the objective attribute characteristics of the history interaction items to form the objective interest sequence of the candidate items comprises: corresponding to each objective interest view angle, acquiring objective similarity between the candidate item and the historical interaction item based on the corresponding item representation; And selecting a plurality of history interaction items from the history interaction items according to the objective similarity corresponding to each objective interest view angle to form an objective interest sequence of the candidate item under the corresponding objective interest view angle.
5. The method according to any one of claims 1 to 4, wherein the processing, by a large language model, the user comment information of the candidate item and the user comment information of the history interaction item, selecting a plurality of items from the history interaction items, and forming a subjective interest sequence of the candidate item includes: Respectively constructing input information with preset prompt information, user comment information of the candidate item and user comment information of the history interaction item, and inputting the input information into a large language model to respectively obtain corresponding scores, pre-emotion characteristics and post-emotion characteristics; Obtaining subjective features of the history interaction item and subjective features of the candidate item according to the history interaction item, scores corresponding to the candidate item, pre-emotion features and post-emotion features; And selecting a plurality of items from the history interaction items according to the subjective characteristics of the history interaction items and the subjective characteristics of the candidate items to form a subjective interest sequence of the candidate items.
6. The method according to claim 5, wherein the obtaining subjective features of the history interaction item and subjective features of the candidate item according to the scores, the pre-emotion features and the post-emotion features corresponding to the history interaction item and the candidate item, respectively, includes: Acquiring a post emotion score of the history interaction item and a post emotion score of the candidate item according to the history interaction item, the scores corresponding to the candidate items and the post emotion characteristics; Acquiring an experience expected gap value of the history interaction item and an experience expected gap value of the candidate item according to the scores, the pre-emotion characteristics and the post-emotion characteristics corresponding to the history interaction item and the candidate item respectively; And constructing subjective features of the historical interaction items based on the post-event emotion scores and the experience expected gap values of the historical interaction items, and constructing the subjective features of the candidate items based on the post-event emotion scores and the experience expected gap values of the candidate items.
7. The method according to any one of claims 1 to 4, wherein the processing, by a large language model, the user comment information of the candidate item and the user comment information of the history interaction item, selecting a plurality of items from the history interaction items, and forming a subjective interest sequence of the candidate item includes: Respectively constructing input information with preset prompt information, user comment information of the candidate item and user comment information of the history interaction item, and inputting the input information into a large language model to respectively obtain a preliminary evaluation category corresponding to the candidate item and the history interaction item, and scores, pre-emotion characteristics and post-emotion characteristics corresponding to the preliminary evaluation category; Clustering and merging the preliminary evaluation category of the history interaction item and the preliminary evaluation category of the candidate item based on semantic similarity to obtain a standardized evaluation category; According to the score, the pre-emotion feature and the post-emotion feature corresponding to each standardized evaluation category of the historical interaction item, subjective features of each standardized evaluation category of the historical interaction item are obtained, and according to the score, the pre-emotion feature and the post-emotion feature corresponding to each standardized evaluation category of the candidate item, subjective features of each standardized evaluation category of the candidate item are obtained; And selecting a plurality of items from the history interaction items according to the subjective characteristics of each standardized evaluation category of the history interaction items and the subjective characteristics of each standardized evaluation category of the candidate items to form a subjective interest sequence of the candidate items.
8. The method according to claim 7, wherein the obtaining subjective features of each standardized rating class of the historical interaction items according to the score, the pre-emotion feature and the post-emotion feature corresponding to each standardized rating class of the historical interaction items comprises: Obtaining the post emotion score of each standardized evaluation category of the history interaction item according to the score and the post emotion characteristic corresponding to each standardized evaluation category of the history interaction item; acquiring an experience expected gap value of each standardized evaluation category of the history interaction item according to the score, the pre-emotion feature and the post-emotion feature corresponding to each standardized evaluation category of the history interaction item; According to the post emotion score and experience expected gap value of each standardized evaluation category of the historical interaction item, constructing subjective features of each standardized evaluation category of the historical interaction item; And/or, according to the score, the pre-emotion feature and the post-emotion feature corresponding to each standardized evaluation category of the candidate item, obtaining subjective features of each standardized evaluation category of the candidate item, including: Obtaining the post emotion score of each standardized evaluation category of the candidate item according to the score and the post emotion characteristic corresponding to each standardized evaluation category of the candidate item; Acquiring an experience expected gap value of each standardized evaluation category of the candidate item according to the score, the pre-emotion feature and the post-emotion feature corresponding to each standardized evaluation category of the candidate item; And constructing subjective features of each standardized evaluation category of the candidate item according to the post emotion score and the experience expected gap value of each standardized evaluation category of the candidate item.
9. The method according to claim 5, wherein selecting a plurality of items from the history interactive items according to the subjective features of the history interactive items and the subjective features of the candidate items to form the subjective interest sequence of the candidate items comprises: based on a preset matching function, calculating subjective matching degree between subjective features of the candidate items and subjective features of each historical interaction item; and selecting a plurality of items from the historical interaction item sequence according to the subjective matching degree to form a subjective interest sequence of the candidate items.
10. The method according to any one of claims 1 to 4, wherein sequentially performing time-series encoding, frequency-domain decomposition and inverse fourier transformation on the objective interest sequence and the subjective interest sequence, respectively, to obtain a corresponding time-domain representation, includes: Respectively carrying out time sequence coding on the objective interest sequence and the subjective interest sequence to obtain corresponding time sequence coding signals; Performing frequency domain decomposition on the time sequence coding signals to obtain frequency domain components of at least two preset frequency bands; and performing inverse Fourier transform on the frequency domain component to obtain a corresponding time domain representation.
11. A user multiple interest recommendation system based on a large model and frequency domain decomposition, the system comprising: The first characteristic acquisition module is used for acquiring objective attribute characteristics of a historical interaction project of a user based on preset project objective attributes; The second characteristic acquisition module is used for acquiring objective attribute characteristics of candidate projects based on the objective attributes of the projects; The objective interest sequence acquisition module is used for selecting a plurality of historical interaction items from the historical interaction items according to the objective attribute characteristics of the candidate items and the objective attribute characteristics of the historical interaction items to form an objective interest sequence of the candidate items; the subjective interest sequence acquisition module is used for processing the user comment information of the candidate item and the user comment information of the history interaction item through a large language model, and selecting a plurality of items from the history interaction item to form a subjective interest sequence of the candidate item; the time domain representation acquisition module is used for sequentially carrying out time sequence coding, frequency domain decomposition and inverse Fourier transform on the objective interest sequence and the subjective interest sequence respectively to obtain corresponding time domain representation; The comprehensive interest representation acquisition module is used for respectively fusing the time domain representations corresponding to the objective interest sequence and the subjective interest sequence to obtain corresponding comprehensive interest representations; and the recommendation module is used for carrying out preference prediction based on the comprehensive interest expression and the objective attribute characteristics of the candidate items and outputting item recommendation results of the users.

Description

User multi-interest recommendation method and system based on large model and frequency domain decomposition Technical Field The invention relates to the field of large model application, in particular to a user multi-interest recommendation method and system based on large models and frequency domain decomposition. Background In the current digital service ecology, the recommendation system is used as a core hub for connecting massive information and user demands, and the performance quality of the recommendation system directly determines the quality of user experience and the commercial value of a platform. To accurately capture the personalized preferences of users, the prior art generally relies on data mining of historical interaction behavior sequences of users, such as clicking, browsing, purchasing, and the like, and describes the evolution process of user interests through a sequence modeling method. To more fully describe the complex interest structure of users in different preference dimensions, a multiple interest modeling approach has developed that attempts to characterize the user's diverse preferences from a parallel perspective by constructing multiple interest vectors. Although the above-described approach improves the accuracy of recommendations to some extent, there are significant limitations in deep mining of user decision psychology and satisfaction driven mechanisms. First, the prior art has limited ability to fuse user subjective experience feedback. The current multi-interest modeling method is mainly based on interactive behavior data, and even if part of schemes introduce user comment texts, the method is generally only remained in general semantic extraction, and is difficult to go deep into specific attribute or aspect dimension. The coarse-granularity analysis mode cannot accurately capture the decision difference of the user at a finer granularity level, so that the depicting capability of the deep mechanism of the interest evolution of the user is insufficient. Secondly, when the existing method processes the historical behavior sequence of the user, the capability of screening key information is insufficient, and the interference of noise interaction is easy to occur. The historical behavior of the user is often disordered and contains a large number of records with low association degree with the current recommendation target. The traditional modeling method still has difficulty in effectively identifying key subsequences truly reflecting the core appeal of the user, so that interest expression is interfered by irrelevant behaviors, and modeling accuracy and result interpretability are affected. Finally, the prior art models the user behavior on a unified time axis, which is easy to cause the model to excessively depend on the historical high-frequency behavior, and also can cause the model to be misled by short-term noise, so that the expression capability of the system on the multi-level dynamic evolution process of the user interest is limited, and the accuracy and individuation level of the recommendation result are further affected. Disclosure of Invention The invention provides a user multi-interest recommending method and system based on a large model and frequency domain decomposition, which are used for improving the matching degree of a recommending result in accuracy and personalized experience. According to a first aspect of the present application, there is provided a user multiple interest recommendation method based on a large model and frequency domain decomposition, the method comprising: based on preset objective attributes of the items, objective attribute characteristics of historical interaction items of the user are obtained; Acquiring objective attribute characteristics of candidate items based on the objective attributes of the items; Selecting a plurality of history interaction items from the history interaction items according to the objective attribute characteristics of the candidate items and the objective attribute characteristics of the history interaction items to form an objective interest sequence of the candidate items; Processing the user comment information of the candidate item and the user comment information of the history interaction item through a large language model, and selecting a plurality of items from the history interaction item to form a subjective interest sequence of the candidate item; Sequentially performing time sequence coding, frequency domain decomposition and inverse Fourier transform on the objective interest sequence and the subjective interest sequence to obtain corresponding time domain representation; Respectively fusing time domain representations corresponding to the objective interest sequence and the subjective interest sequence to obtain corresponding comprehensive interest representations; and carrying out preference prediction based on the comprehensive interest expression and the objective attribute