Recommendation model based on double-layer self-attention comment modeling

Abstract

The invention discloses a recommendation model based on double-layer self-attention comment modeling. The model comprises a user portrait module, an article portrait module and an interaction module.The user portrait module and the article portrait module are the same in structure, firstly, related words spaced by any distance in sentences are flexibly combined by introducing self-attention in aphrase extraction layer, and article feature phrases and emotion phrases are formed; then article feature phrases are associated with emotion phrases by using self-attention in the phrase associationlayer to obtain emotion polarity of a user for each article feature for constructing a user-article portrait, and finally experimental verification is performed on the model on six data sets from Amazon 5-core. According to the method, the self-attention network is introduced into comment modeling of the recommendation system, the emotional polarity of the user to the'article features' is considered under the deep learning framework, the problems of noise and context loss caused by phrase extraction of the CNN are relieved, the user-article portrait is modeled in a fine-grained mode, and the recommendation performance is improved.

Description

technical field [0001] The invention belongs to the field of recommendation systems, in particular to a recommendation model based on double-layer self-attention comment modeling. Background technique [0002] In recommender systems, traditional collaborative filtering methods, which infer user and item behavior patterns from rating data, are still competitive techniques today. However, when encountering data sparsity and cold start problems, the performance of collaborative filtering methods has dropped significantly. Therefore, many researchers have tried to introduce review texts as a supplement, mining multiple elements in reviews to model user-item portraits. The seven common elements in reviews are "high-frequency words", "review topic", "sentiment for item features", "contextual sentiment", "contrastive sentiment" and "emotion", these elements have been obtained in the traditional machine learning era Deep research. [0003] However, existing deep learning methods o...

AI Technical Summary

Problems solved by technology

This model uses the self-attention network to mine the user's emotions on different aspects of the item, and constructs a fine-grained user-item portrait; and solves the problem of variable-length phrase extraction, alleviating the problems of noise and context loss introduced by CNN's fixed window

Images