Cold Start Recommendation Algorithm Based on Latent Factor Prediction

Abstract

At present, recommendation algorithms have been widely used in online shopping, commercial services and entertainment. Although the recommended methods are endless, they are generally recommended based on the user's historical behavior or project information. Since the emergence of collaborative filtering in the 20th century, it become a very popular method in recommendation. But on the issue of cold start, collaborative filtering is not effective due to the lack of information. In this paper, we focus on solving the user side cold start problem. We propose a cold start recommendation algorithm based on latent factor prediction by combining collaborative filtering algorithm and matrix factorization. When it is just an ordinary user, we predict the rating by matrix factorization, and during the training iteration, we propose an improved optimization algorithm based on Alternate Least Squares (ALS). When the user is a new user, we predicted the latent factor by feature prediction. And then, the final outcome is calculated. The experiment uses the movielens data set. The experiment results show that our proposed recommendation algorithm has approximately 20.4% improvement over the baseline cold start algorithm. In addition, this paper design experiment to validate the proposed optimization algorithm effect, and the experiment results prove that the proposed algorithm has good precision and scalability.