Boosting the K-Nearest-Neighborhood based incremental collaborative filtering

Abstract

Recommender systems which can automatically match users with their potential favorites usually rely on Collaborative Filtering (CF). Since in real-world applications the data of historical user behavior are ever growing, it is important to study the incremental CF models which can adapt to this data explosion quickly and flexibly. The rating similarity based K-Nearest-Neighborhood (RS-KNN) is a classical but still popular approach to CF; therefore, to investigate the RS-KNN based incremental CF is significant. However, current incremental RS-KNN (I-KNN) models have the drawbacks of high storage complexity and relatively low prediction accuracy. In this work, we intend to boost the RS-KNN based incremental CF. We focus on two points which are respectively (a) reducing the storage complexity while maintaining the prediction accuracy by employing the generalized Dice coefficients, and (b) improving the prediction accuracy by integrating the similarity support and linear biases as well as implementing the corresponding incremental update. The efficiency of our strategies is supported by the positive results of the experiments conducted on two real datasets.