Using Stochastic Gradient Decent Algorithm For Incremental Matrix Factorization In Recommendation System

In recent years, with the development and daily update of user feedback data, the recommendation systems require algorithms able to process data accurately and fast. To tackle this data stream problem, online model updates for new data points come to be available. Another issue is how to evaluate algorithms in a streaming data environment while conventional Collaborative Filtering algorithms are proposed for stationary data. Furthermore, traditional evaluation methodologies are only useful in offline experiments. In this research, we propose a novel incremental model base on stochastic gradient decent algorithm for Matrix Factorization. In addition, we also show a prequential evaluation protocol for recommender systems, applicable for streaming data environments. Comparing with other state-of-the-art models, our algorithm has combative accuracy, while being significantly faster.