Research On Visual Feature Based Recommendation Systems

Data sparsity problem has long been a big challenge in recommender system without perfect/good solution proposed so far.In practical recommender system,the performance of traditional recommendation algorithms is highly limited due to sparse data.Therefore,solving rating data sparsity problem is of great importance to the performance improvement of recommender system.As for movie recommender system,however,visual features would bring much more valuable information due to the fact that most of the user feedbacks greatly come from movies' visual features,which therefore could help movie recommender system make more accurate and reliable recommendation.Visual information is a kind of essential information for human to understand the world.The abundant information of visual features absolutely give people useful inferences.The less attention is possibly because of the difficulties of extracting abstract features from images and movies via traditional feature extracting methods.Under the circumstance,the movie recommendation methods is deeply investigated in this thesis.The main contributions are as follows:Firstly,a review of recommendation techniques is given,including the development of recommendation,movie recommendation methods and visual feature extracting schemes.Secondly,a visual feature based movie recommendation system architecture is proposed.There are three modules in the system:data collecting module,visual feature extracting module and recommendation module.The most important and distinct module is,visual feature extracting module,which largely influence the performance of two proposed methods.Thirdly,a visual feature based movie recommendation model is proposed to improve the accuracy.The visual features extracted from videos are used as auxiliary information to help ratings to complete recommendation process.Therefore,the model contains two essential parts,one is visual features extracting,and another is movie recommendation generating.The method of extracting visual features from trailers are:1)electing output of a Convolutional Neural Network(CNN)layer as visual feature vector,2)utilizing video caption and document(label)analysis,generally,firstly using video caption to label videos and then extract textual features from labels.In movie recommendation part,Probabilistic Matrix Factorization(PMF)is selected to generate movie recommendation,because of its feasibility and high-level performance.Large amounts of experiments show the good performance of the proposed model.Fourthly,considering the rich information included in document,such as reviews,textual information is selected to collaborate with visual information helping ratings together.The implementation schemes of the proposed model are:1)extracting visual features and textual feature from videos and reviews,respectively,then performing PMF method on two vectors simultaneously,2)firstly utilizing video caption to label videos,then putting labels and reviews together,and extracting features from the document,then performing PMF method on the extracted feature vector.Amounts of experiments demonstrate the good performance of the proposed model.Finally,the summary of the overall thesis is given.Then the contribution and future work are analyzed.