CN-121980091-A - Mixed identifier generation type recommendation method and system based on local cooperative context

CN121980091ACN 121980091 ACN121980091 ACN 121980091ACN-121980091-A

Abstract

The invention belongs to the technical field of artificial intelligence, in particular to a hybrid identifier generation type recommendation method and a system based on local cooperative context, which are used for acquiring a user history interaction sequence and carrying out multi-granularity clustering on users to obtain groups of each user; the method comprises the steps of establishing a static identifier based on article content, establishing a dynamic identifier by combining user group information and a local interaction sequence, fusing to generate a mixed identifier of each article, establishing an instruction fine tuning task, embedding the group information into an instruction prompt, learning a history sequence to generate a mixed identifier of the next article by using a large language model, and generating a recommendation result based on the optimized large language model according to the history sequence of a target user and the group information of the target user. According to the invention, the local cooperative context is fused through multi-granularity group estimation, and the mixed article representation is constructed by combining the static identifier and the dynamic identifier, so that the problems of single user interest modeling and semantic loss of article representation are effectively solved, and the accuracy and individuation level of the generated recommendation are improved.

Inventors

CUI XU
LIU JINHUAN
MA SHUHUI
SONG XUEMENG
YU YANWEI
DU JUNWEI

Assignees

青岛科技大学

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. A hybrid identifier generation type recommendation method based on a local cooperative context, comprising the steps of: Coarse-granularity clustering is carried out on the long-term preference representation, fine-granularity clustering is carried out on the short-term preference representation, and a group finally distributed by each user is obtained; The method comprises the steps of establishing a static identifier based on content data of articles in a historical interaction sequence, extracting a local interaction sequence according to the historical interaction sequence of a user, and establishing a dynamic identifier by combining group information of the user; constructing an instruction fine-tuning task, embedding group information finally distributed by a user into an instruction prompt, taking a user history sequence as input and taking a mixed identifier of the next article as target output; and generating a recommendation result based on the optimized large language model according to the historical interaction sequence of the target user and the group information thereof.
2. The hybrid identifier generation type recommendation method based on local cooperative context according to claim 1, wherein the long-term preference representation and the short-term preference representation of the user are extracted, specifically: encoding the complete historical interaction sequence of the user through a pre-training BERT model to obtain long-term preference expression; and (3) coding the local interaction sequence of the last d articles of the user through a pre-training BERT model to obtain short-term preference expression.
3. The method for generating a recommendation based on a hybrid identifier of a local collaborative context according to claim 1, wherein coarse granularity clustering is performed on long-term preference representations, fine granularity clustering is performed on short-term preference representations, specifically, clustering is performed on long-term preference representations and short-term preference representations respectively to obtain coarse granularity groups and fine granularity groups, and the coarse granularity clustering result is used as a constraint, and the fine granularity clustering result is iteratively optimized through a reward function with enhanced time sequence stability and an Actor-Critic framework to obtain groups finally allocated by each user.
4. The method for generating a recommendation based on a hybrid identifier of a local collaborative context according to claim 3, wherein the method further comprises iteratively optimizing a fine-grained clustering result, pre-optimizing the fine-grained clustering center through a variance-aware adaptive learning rate mechanism before policy optimization, and adjusting the learning rate according to a variance of a reward variance among users to improve stability of cluster center updating.
5. The hybrid identifier generation-based recommendation method based on local collaborative contexts of claim 1 wherein static identifiers are constructed, specifically wherein the content of the item is multi-level quantized by a pre-trained residual quantization variation self-encoder, a set of discrete codes is generated from a plurality of codebooks as static identifiers, and codebook parameters are optimized by a quantization loss function such that the static identifiers preserve core semantic information of the item.
6. The method for generating a recommendation of a hybrid identifier based on local collaborative contexts as claimed in claim 1, wherein the method is characterized by constructing a dynamic identifier, specifically, extracting a local interaction sequence of a corresponding user according to a historical interaction sequence of the user, determining d items recently interacted by the user, combining with group information of the user, guiding a large language model to generate a semantic tag of the dynamic context for the d items recently, constructing the dynamic identifier, and constructing the dynamic identifier by taking a residual error coded by the static identifier as a starting point, and selecting codewords step by step through residual error quantization.
7. The hybrid identifier generation based on local collaborative context recommendation method of claim 6, wherein building dynamic identifiers further comprises bringing items with the same static identifier closer to each other in semantic space and items with different static identifiers farther from each other by a diversity regularization loss function, ensuring semantic discrimination of dynamic identifiers.
8. The method for generating a hybrid identifier recommendation based on a local collaborative context according to claim 1, wherein generating a hybrid identifier for each item further comprises aggregating each level of encoded embedment of a static identifier with each level of encoded embedment of a dynamic identifier to obtain quantized embedments, reconstructing semantic information by a decoder to construct a semantic reconstruction loss function to ensure semantic fidelity of the hybrid identifier.
9. The hybrid identifier generation-based recommendation method of claim 1, wherein the constructed instruction hints-tuning tasks include an identifier-to-identifier symmetric prediction task, an identifier-to-text asymmetric prediction task, and a text-to-identifier asymmetric prediction task; Inputting a mixed identifier sequence of a user history item for a symmetric prediction task from identifier to identifier, and outputting a mixed identifier of a next item; inputting a mixed identifier sequence of a user history item for an asymmetric prediction task from an identifier to a text, and outputting a text description of the next item; for the text-to-identifier asymmetric prediction task, a text description of the user's historical items is entered and a hybrid identifier of the next item is output.
10. A hybrid identifier generation recommendation system based on a local collaboration context, comprising: The multi-granularity group estimation module is configured to acquire a historical interaction sequence of a user, extract a long-term preference representation and a short-term preference representation of the user, perform coarse granularity clustering on the long-term preference representation, and perform fine granularity clustering on the short-term preference representation to obtain a group finally distributed by each user; the mixed identifier construction module is configured to construct a static identifier based on content data of the articles in the historical interaction sequence, extract a local interaction sequence according to the historical interaction sequence of the user, and construct a dynamic identifier in combination with group information of the user; the large language model fine tuning module is configured to construct an instruction fine tuning task, embed group information finally distributed by a user into an instruction prompt, take a user history sequence as input and take a mixed identifier of a next article as target output; And the online recommendation module is configured to generate a recommendation result based on the optimized large language model according to the historical interaction sequence of the target user and the group information thereof.

Description

Mixed identifier generation type recommendation method and system based on local cooperative context Technical Field The invention belongs to the technical field of artificial intelligence, and particularly relates to a hybrid identifier generation type recommendation method and system based on local cooperative contexts. Background The statements in this section merely mention background of the present disclosure and do not necessarily constitute prior art. The generated recommender is a research hotspot in the area of intersection of artificial intelligence and recommender, and aims to generate representations of items of potential interest to users through a deep learning model, rather than simple screening from a predefined candidate set. Currently, researchers try to apply large language models to generative recommendation tasks, and the recommendation performance is improved by using the powerful semantic understanding capability and knowledge migration capability of the large language models. However, current methods rely primarily on the historical interaction sequences of individual users and global user identifiers (ID information) for preference learning, forcing the model to infer user interests only from the global collaborative signal. This modeling approach ignores the diversity and dynamic nature of the user interests. In practical applications, users often have interest preferences in multiple dimensions at the same time, and the interests may exhibit different characteristics on a time scale, for example, long-term interests are relatively stable and suitable for large-scale user grouping, short-term interests reflect recent behavior trends, and more precise alignment between users with similar new interests can be achieved. The existing method cannot effectively fuse user preference information with different granularities, so that negative migration problems occur when the model faces users with diversified interests, namely, interactive data of the users with obvious interest differences are mixed indiscriminately, and recommendation accuracy is lowered. Disclosure of Invention The invention provides a local collaborative context-based hybrid identifier generation type recommendation method and system, which are characterized in that long-term and short-term preferences of users are captured through multi-granularity user group estimation, static semantic stability and dynamic context perceptibility are fused by utilizing residual hybrid identifiers, and group collaborative information is injected into a large language model through context-guided instruction fine tuning. The method comprises the steps of finely dividing users, constructing enhanced article representations, and finally enabling a model to learn recommendation under group guidance, so that accuracy and robustness of generated recommendation are improved. The first aspect of the invention discloses a hybrid identifier generation recommendation method based on local cooperative context, comprising the following steps: Coarse-granularity clustering is carried out on the long-term preference representation, fine-granularity clustering is carried out on the short-term preference representation, and a group finally distributed by each user is obtained; The method comprises the steps of establishing a static identifier based on content data of articles in a historical interaction sequence, extracting a local interaction sequence according to the historical interaction sequence of a user, and establishing a dynamic identifier by combining group information of the user; constructing an instruction fine-tuning task, embedding group information finally distributed by a user into an instruction prompt, taking a user history sequence as input and taking a mixed identifier of the next article as target output; and generating a recommendation result based on the optimized large language model according to the historical interaction sequence of the target user and the group information thereof. Further, a long-term preference representation and a short-term preference representation of the user are extracted, specifically, the whole historical interaction sequence of the user is encoded through a pre-training BERT model to obtain the long-term preference representation, and the local interaction sequence of the last d articles of the user is encoded through the pre-training BERT model to obtain the short-term preference representation. Further, coarse granularity clustering is carried out on the long-term preference representation, fine granularity clustering is carried out on the short-term preference representation, specifically, the long-term preference representation and the short-term preference representation are clustered to obtain a coarse granularity group and a fine granularity group, and the coarse granularity clustering result is used as a constraint, and the fine granularity clustering result is iteratively optimized throug