Multiple metadata based collaborative filtering using random walk on a movie K-partite graph for movie recommendation

Abstract

A recommendation system is an application to recommend items or information which interest to a user by analyzing the user`s preference. Many researchers have developed various recommendation techniques to provide high quality recommendation to users. Collaborative filtering which recommends items to a user based on interest of other users having similar preferences is known to be the most popular approach in the recommendation system. However, sparse and high dimensional input data for collaborative filtering cause poor recommendation performance. I thus utilize external meta information to make a dense form of input data for collaborative filtering. In this paper, I propose a novel Multiple Metadata based Collaborative Filtering (MMCF) and use such multiple metadata as genre, director, and actor. In detail, MMCF builds a movie k-partite graph of users, movies and multiple metadata, and extract their implicit relationships between metadata and between users and metadata. Then it propagates the implicit relationships between metadata into the implicit relationships between users and metadata by applying random walk process in order to alleviate sparsity of the original data. Finally, I empirically evaluate the accuracy of my MMCF on the real Netflix movie dataset, and achieve 10.5% RMSE improvement over previous collaborative filtering methods.