Research For Recommender Systems Based On Deep Matrix Factorization

In the information age,the massive information generation leads to information explosion,which results in information overload.This is a dilemma that people are facing.The recommendation system is a powerful tool to alleviate this situation by sat-isfy a user through personalized recommendation.Which means that the information produced by producer can be quickly delivered to the consumer to achieve a win-win situation.Recommendation system plays an important role in the information tech-nology industry such as e-commerce,information distribution,music industry and so on.Traditional recommendation algorithms can be divided into collaborative filtering(CF),content-based recommendation algorithm and hybrid recommendation algorithm.CF utilizes the interaction history between users and items to recommend new items to the users,matrix factorization has always been the forefront of CF research.Content-based recommendation is to recommend user some items which are kind of similar with those items that the user has already interacted.so one of the advantage of such algo-rithm is that we can explain it,it's more convincing.Hybrid recommendation algorithm is to integrate multi-source heterogeneous information to make recommendation.This paper focuses on the application of neural network in the recommendation system,especially on matrix factorization.Firstly,the recommendation system,col-laborative filtering and matrix factorization are briefly introduced.Then introduce the background work of this paper,such as neural collaborative filtering,deep structured semantic model and so on.Based on these basic models and methods,this paper expand from several aspects.Firstly,deep matrix factorization model is proposed,according to the user's histor-ical rating information,a user-item interaction matrix is constructed with both explicit ratings and non-preference implicit feedback,then use neural network to map the users and items into a common deep structured semantic space with non-linear projections.Secondly,we design a new loss function based on cross entropy,which considers the explicit ratings.Finally,the user's state should be different depending on the item of the interaction,because the item focus on different part of the user's interaction history.So we introduce the attention mechanism into deep matrix factorization model and pro-pose an attention based deep matrix factorization model which can further improve the result.