Research On End-to-End Movie Recommender System Exploiting Visual Content

In the mobile internet era,services of personalized recommendation have facilitated the daily life of human beings.The growing data has become a label in the internet era and has become one of the most valuable resources.However,a large number of redundant data are not effectively utilized in recommender systems,and the high sparsity of the interaction data between the user and the item brings a lot of troubles to the personalized recommendation,and also brings great challenges to the research work.In movie recommendation,a few research work consider the visual contents of movies,and the aesthetic quality assessment of movie posters and still frames have not been used in movie recommendation.Also,the existing works use deep learning to extract the visual contents of images and take them into recommender system.However,these visual features are only designed for classification.The learning of visual features are separated from recommendation,and the performance of personalized recommendation has not improved obviously.This paper combines deep learning and recommendation algorithms to integrate visual and aesthetic content into the End-to-End movie recommender system,to alleviate the high sparsity of data and improve the performance of personalized movie recommendation.The main contents of this paper are as follows:(1)We analyze the research background and research state both domestically and abroad of current recommender systems,then we discuss the advantages and disadvantages of tra-ditional recommendation algorithms and the development trend of movie recommendation.(2)We analyze and utilize the visual content contained in movie posters and still frames.In this paper,we propose an end-to-end movie recommendation model exploiting visual fea-tures.Specially,we use the deep convolutional neural network to extract the visual features of the image.And the parameter learning of visual features and recommendation are inte-grated into a unified framework for training.We conduct experiments on two real datasets and compare with advanced methods.The results demonstrate the feasibility and effective-ness of the proposed method.(3)We analyze the effect of aesthetic features contained in movie posters and stil-1 frames on personalized recommendation.In this paper,we propose an end-to-end movie recommendation model exploiting aesthetic features.Specially,we use the deep aesthetic model to extract the aesthetic features designed for aesthetic quality assessment.Further-more,the parameter learning of aesthetic features and recommendation are integrated into a unified framework for training.We conduct experiments on two real datasets and com-pare with advanced methods.The results demonstrate the feasibility and effectiveness of the proposed method.(4)We develop a prototype system of movie recommendation that combines visual features and aesthetic features based on end-to-end.