Research And Implementation Of Collaborative Heterogeneous Information Embedding For Recommender Systems

The rapid improvement of artificial intelligence(AI)technology has brought opportunities to many fields and recommender system is one of the typical fields in which AI technology has been successfully applied.Collaborative filtering(CF)has emerged as one of the fastest growing and most widely used algorithms in the history of recommender systems since it was proposed in the Netflix Prize.Various AI algorithms have achieved certain success in the application of CF,however,sparsity of user-item interactions and cold start problems degrade the performance of improved CF algorithms significantly.Using auxiliary information is a common way to alleviate these two problems,but previous methods deal with different types of information separately,which inevitably results in the loss of effective information.Heterogeneous information networks(HINs)are good at processing multiple node types and complex edge types between nodes.Therefore,this thesis do recommendation tasks in HINs and integrates heterogeneous information to enhance the performance of CF methods.The main purpose of this thesis is to make full use of the rich and effective information contained in HINs as well as the excellent feature learning and inference ability of Bayesian deep learning methods.A hybrid recommendation method based on collaborative heterogeneous information embedding is proposed to enhance traditional CF.In the implementation process,the recommender system based on HINs needs to face two problems.One is how to effectively represent the high-level semantic information in the network,and the other is how to integrate the heterogeneous information to enhance the performance of the downstream recommendation task.In order to solve these two problems,this thesis uses meta-path to capture the rich semantic information contained in HINs and use deep learning to learn the high-level feature representation of nodes in HINs.At the same time,it uses "PMF + Deep Learning" framework to achieve heterogeneous information embedding and get the potential low-dimensional feature representation vectors of users and items.The product of two vectors is the final predicted rating,and the N items with the highest score of a user can be recommended to the corresponding user.In this thesis,We conduct experiments on a real movie recommendation network named extended MovieLens dataset,which show that our proposed approach outperforms the state-of-the-art recommendation techniques in MAE and RMSE.Moreover,the heterogeneous information used in this thesis contains the properties information of the movie,which can alleviate the cold start problem.In addition,this algorithm is performed in HIN,which increase the interpretability of recommendation method.