Research Of Recommendation Technology Based On Deep Representation Learning

In the big data era,it is difficult for people to find valuable information from the explosively growing data.A recommender system has been proposed to solve the problem and widely used in our life in recent years.As users and items are indispensable objects to a recommender system,user and item representation learning plays an important role for recommendation methods.Generally,recommendation methods use users' historical behavior data on items to predict what they may like.The challenge is how to provide a powerful method to represent users and items from such sparse historical behavior data.We can solve the problem by combing multiple methods and integrating multiple data sources.On the one hand,the strategy of combining multiple methods can be used to represent the latent features of users and items from different perspectives,which can reduce the impact of sparse data on recommendation quality to a certain extent;on the other hand,the strategy of integrating multiple data sources can provide richer feature information of users and items to improve recommendation quality.Based on the above two observations,this paper employs the deep learning methods to learn latent representations of users and items for recommendations.The main work as follows:(1)This paper proposes a new top-N recommendation method to learn the latent features of users and items by combining the Bayesian Personalized Ranking(BPR)and Collaborative Less-is-More Filtering(CLi MF).This method optimizes the combined heterogeneous loss function and learns the latent features of users and items to make recommendations.The experimental results show that the combined loss function enforces structure diversity of user pair preference,which makes it improve the top-N recommendation quality.(2)This paper proposes a recommendation method of integrating users' social information to learn the latent features of users and items.This method uses a deep end-to-end framework that utilizes trust network embedding tasks to assist the recommendation task.The two tasks are associated with a cross unit,which automatically combines latent features of users both in the recommender system and the trust network.The experimental results on two real datasets show that this method improves the performance of the rating prediction task.(3)This paper proposes a recommendation method of integrating users' review information to learn the features of users and items.In this method,the adversarial learning strategy among the relationship of rating and review is used to learn auto-encoders.The learned auto-encoders can be used to extract the latent features of users and items for recommendations.The empirical studies on real-world datasets show that the proposed method improves recommendation performance.