With the help of recommender systems, users can find information they do not know about, but they are interested in. These systems are a specific type of intelligent systems that track each user's behavior and discover their behavioral patterns. Among different approaches for implementing recommender systems, Matrix-Factorization (MF) based methods are so popular due to their high accuracy and scalability. However, in real world applications that new ratings are continuously coming, processing the huge amount of data is a computationally expensive task. In this paper, we present a novel incremental matrix factorization method to learn only parts of the data that have been changed or added recently. This way, there is no need to train the system from the scratch. The input data to the proposed recommender system is of two types, batch data and stream data. Batch data is the rating data that is already saved in the system and relates to the activities of users in the past. Stream data is the rating data that enters the system in each time interval. The method is evaluated on two versions of popular MovieLens dataset from GroupLens research. The experimental results confirm that our method reduces the execution time significantly while keeping the prediction error intact.
展开▼
