SWCB: An Efficient Switch-Clustering of Bandit Model

Abstract

Bandit model is a general framework for solving the cold-start problem in recommender systems. In recent works, clustering strategies have been adopted in the bandit settings to further mitigate the lack of information for new users and items. Although these approaches prove to be effective, their performance may suffer from the clustering efficiency. On one hand, quickly identifying the underlying cluster to share the information has a significantly impact on retaining new users with low levels of interaction in the early stages of recommendation. On the other hand, problems like overfitting might result from repeatedly using the same clustering algorithm on the same set of users, which may decrease the recommended performance over an extended period in the later stages. To address these issues, we propose a Switch-Clustering of Bandit (SWCB), which utilizes the Bandit framework to enable online recommendation as well as new user clustering in the cold-start phase. In particular, the proposed model improves the efficiency of recommendations by switching the clustering algorithms automatically. SWCB can quickly identify the underlying clusters by splitting and merging operation. When clustering metric updated by Bandit reaches a certain threshold, SWCB alters the way of user clustering, thus speeding the information retrieval in the cold start stage and helping to retain new users of online recommendations. We evaluate the efficiency and effectiveness of the proposed model on one synthetic dataset and two real-world datasets. Compared with the baseline bandit models, SWCB gets higher rewards within the rounds of the cold-start phase.