Predict New Users' Taste by Modeling Users' Latent Features

100

Recommendation system is popular in recent years. A key challenge in recommendation system is how to characterize new users taste effectively. The problem is generally known as the cold-start problem. New users judge the system by the ability to immediately provide what they are interesting. A general method for solving the cold-start problem is eliciting new users’ information by answering interview questions. In this paper, we present Matrix Factorization K-Means (MFK), a novel method to solve the problem of interview question construction. MFK first learns the user and item latent features by the observed rating data and then determine the best interview questions based on the clusters of latent features. We can find which group of users they are similar to after attain responses of the interview questions. Systems can indicate the new users’ taste according to their response on interview questions. In our experiments, we evaluate our methods in a public dataset for recommendation. The results show our method leads to a better performance compared with other baselines. Besides, the performance of our method is close to the state-of-the-art technique, while our method has a better computation complexity.