Video recommendation based on multi-modal information and multiple kernel

Abstract

Collaborative Filter (CF) algorithms often suffer from data sparsity and item cold start problem, for the user-item matrix is insufficient and extremely sparse especially when new item is added to recommendation system. These two problems also exist in video recommendation process. We propose two methods to solve them by incorporating multimodal information and multiple kernel together. To solve item cold start problem, we learn a user taste hyper-plane by using multiple kernel SVM to represent the user taste, which is further used to predict the recommendation of new added videos. We combine the different user taste hyper-plane similarity and the traditional cosine similarity with a trade-off between them to address the data sparse problem. Experimental results show that our proposed algorithm can alleviate the data sparsity and item cold start problems.