Learning for top-N recommendations High-dimensional and heterogeneous information

Top-N recommendations have been widely adopted to recommend ranked lists of items to help users identify the items that best fit their tastes. Various efforts have been dedicated to providing top-N recommendations by learning effective models from data. A recent trend is to utilize information in the multimedia scenario, where the data are generally of high-dimensionality and are constructed from multiple sources of information. In this thesis, we provide top-N recommendations by learning from high-dimensional and heterogeneous information. Specifically, we designed a new regularization for item-based models; we studied embedded feature reduction techniques to utilize high-dimensional side information; to exploit noisy information, we devised a new network structure based on variational auto-encoder; we proposed a graphical model to combine cold-start rating information and side information to recommend new items; we provided top-N recommendations to generate a list of candidates for paper reranking; and we investigated personalized feature interaction selection for factorization machines.