Research On Recommender System Based On Deep Multiple Matrix Factorization

With the rapid development of e-commerce,the users in the face of the diversifi-cation of products and services,how to quickly and efficiently find suits theirs own items,that become a major concern of the IT companies.Recommender system aims to help users to save time and effort cost by selecting items,and tries to make accurate personalized recommended for users.An excellent recommender system can help users to find items what they really need more easily.It also can effectively help enterprises to keep the users and improve potential users' conversion rate.In recent years,deep learning has made remarkable achievements in computer vi-sion,speech recognition and natural language processing.At the same time,the re-searches and applications in the aspect of recommender system have been more and more known.Meanwhile,the method named matrix factorization is used to do research in recommender systems.Matrix factorization is used to predict that if a particular user likes an item or not,according to the similarity between users and items.In this thesis,we adopt the following methods:first,we set up rating matrix ac-cording to the users' ratings of individual items.Secondly,we use the rating matrix as input.Using ratings of a certain user to all items and all ratings of the certain item caused by all users as the inputs of neural network respectively.Each layer of neural network gets two different vectors,calculating the cosine similarity between the two vectors,and all of the cosine similarity results are added up to get the final results.Thirdly,we use the modified binary cross entropy loss function as loss function of neu-ral network's training.Through the modified binary cross entropy loss function,we can effectively use the display ratings and implicit information to optimize the model.Compared with the outstanding algorithms,our experiment's results also show that our model performs better than others.At the same time,a sensitivity analysis was conducted for different neural network parameters,such as network layers,negative sampling rate.