Personalized Recommender System Research Based On Deep Learning

With the development of Internet and information technology,Recommender System is a more and more important research topic.There are two typical types of recommender system.Content-based and collaborative filtering-based.Deep learning has achieved tremendous success in computer vision and natural language processing.Its feature extraction power attracts great attention to recommender system researchers.In deep-learning-based collaborative filtering models,one good type of model is autoencoder-based collaborative filtering model.But existing implementation of this model didn't exploit the sparsity structure of the data,making training slow,impractical to today's big data environment.Furthermore,this model only uses the rating information,it performs worse when in the cold-start scenario.To tackle above problems,our key contributions as follows:1.We briefly introduced the key concept of Content-based recommender system and Collaborative-Filtering based recommender system.In terms of Collaborative-Filtering based recommender system,we analyzed two main popular approaches,memory-based and model-based.2.We introduced our key deep learning based collaborative filtering model:the autoencoder based collaborative filtering model.To tackle the data sparsity issue and to speed up training.We exploited the sparsity pattern of training data and proposed a speed-up training method.Experiments show orders of magnitude speed up with comparable accuracy comparing to existing autoencoder-based CF method.3.Based on previous works,we analyzed the multi-layer deep autoencoder and updated the greedy layer-wise training scheme to fit the rating data.Our method solves the gradient vanishing problem.Experiments show our multi-layer deep autoencoder model greatly outperforms two-layer model,with a moderate time cost.4.To solve the cold-start problem.We proposed a method to integrate side information into our network.Experiments show that in cold start scenario,adding side information greatly improves recommendation accuracy.