Research On Explainable Recommender System Based On Multi-Task Learning

With the rapid development of all walks of life in the Internet era,the online process is changing rapidly,and a large number of Internet users are producing a large number of complex data all the time,in order to enable users to get the content they are interested in,recommender system emerges as the times require and is widely used in various fields of Internet applications.With the development of explainable artificial intelligence,explainable recommender system has gradually become a new research hotspot.A highly explainable recommender system not only recommends items to users,but also provides recommendation reasons.Recommendation reasons help users better understand the recommendation results and provide important reference for users to make decisions.This paper focuses on mining the information of users and items in the reviews text for explainable recommendation.However,the existing explainable recommender systems do not consider the differences of user's and item's reviews,and the portraits of users and items are not accurate enough,which leads to inaccurate rating prediction and generate recommendation reason text with insufficient personalization and poor content.In order to solve the above problems,the main research contents of this paper are as follows:(1)In the problem of recommendation reason generation,in order to generate highquality text with strong personalization and rich content,this paper proposes an item recommendation reason generation model based on Transformer model,which fuses aspect sentiment and external knowledge.The model uses Bi-Directional Attention Flow network twice to fuse the title of the item,item's aspect sentiment what users like and external knowledge,so as to generate high-quality text.In the process of obtaining item's aspect sentiment what users like,this paper also proposes a method of twice training BERT model by fusing domain knowledge to do aspect-based sentiment analysis of reviews text.The experimental results on real public datasets show that the BERT-DK model is effective for aspect-based sentiment analysis,and the recommendation reason generation model can generate personalized and rich-content text for users.(2)Due to the differences between user's and item's reviews,this paper proposes an improved asymmetric attention hierarchy model to analyze the reviews and accurately represent the user and item.Based on the accurate representation of users and items as the sharing module,this paper proposes a multi-task learning explainable recommendation model considering the correlation between the rating prediction task and the recommendation reason generation task.The model trains two subtasks jointly,Trans FM model is used to complete the rating prediction task,and the recommendation reason generation model proposed in this paper is used to complete the recommendation reason generation task.Experimental results on several real public datasets show that the proposed multi-task learning explainable recommendation model has higher accuracy and better explainability than other similar recommender model.There are 21 figures,15 tables,and 84 references in this paper.