A Hybrid Collaborative Filtering Model With Deep Learning And Social Regularization For Recommendation

With the rapid development of the Internet,network information shows explosive growth and its structure is also becoming increasing complex.Due to the huge amount of information,it's very difficult for customers to find the content they are interested in.Recommender systems can further explore the deep-seated user demand and provide personalized experience to users.In addition,recommender systems can help customers to find out what they need more easily, and improve loyalty of their customers by improving user experience, then further convert more potential customers into consumers.Besides,it is worth studying in the field of the computational mathematics,cognition science and information science.Collaborative filtering(CF) is a widely used approach in recommender systems. Traditional CF-based methods employ the rating matrix for recommendation.The rating matrix is usually very sparse.Due to this sparsity problem,traditional CF will suffer from unsatisfactory performance.In this case,some models utilize the auxiliary information to address the data sparsity problemas well as the cold start problem.However,the learned feature representation may not be effective when the auxiliary information is very sparse.To address this problem, we propose a novel mode,called A Hybrid Collaborative Filtering Model with Deep Learning and Social Regularization.We use Social Regularization to represent the social constraints on recommender systems.Our model utilize the advances of learning effective representations in deep learning and combine deep item's latent factors vector learning from auxiliary information with rating matrix.In addition, social regularization terms are used to constrain the objective function so that relevant items share a high similarity which leads to further improve recommendation performance. CDL-SR could also provide information recommendation for new users(cold start) and address the items text content,property as well as streams sparsity problem.Experiments on real CiteUlike data show that our hybrid model outperforms other methods in effectively combine deep learning with social networks information and achieves performance improvement.Especially we compare the the CDL-SR with CTR,CDL-SR has an improvement of 66.7% for recall.Moreover,the recommendation system using CDL-SR can make good recommendation and the convincing reasons which are helpful for further improving customer satisfaction.