Research Article
Clustered UCB-C: Exploiting User Heterogeneity for Low-Regret Recommendation in Structured Multi-Armed Bandit Settings
@INPROCEEDINGS{10.4108/eai.22-5-2026.2365117, author={Zhitao Wang}, title={Clustered UCB-C: Exploiting User Heterogeneity for Low-Regret Recommendation in Structured Multi-Armed Bandit Settings}, proceedings={Proceedings of the 4th International Conference on Image, Algorithms, and Artificial Intelligence, ICIAAI 2026, 22-24 May 2026, Singapore, Singapore}, publisher={EAI}, proceedings_a={ICIAAI}, year={2026}, month={8}, keywords={UCB-C algorithms heterogeneity cumulative regret recommendation systems}, doi={10.4108/eai.22-5-2026.2365117} }- Zhitao Wang
Year: 2026
Clustered UCB-C: Exploiting User Heterogeneity for Low-Regret Recommendation in Structured Multi-Armed Bandit Settings
ICIAAI
EAI
DOI: 10.4108/eai.22-5-2026.2365117
Abstract
The heterogeneity of user preferences is a core challenge in recommendation systems. Traditional MAB algorithms have limitations such as single clustering and fixed α. Therefore, it is crucial to enhance their practicality in recommendation systems. Based on the MovieLens 1M dataset, this paper proposes a two-step strategy of "preference clustering and users clustering" and personalized α allocation, and designs the clustering UCB-C algorithm to study low regret recommendations. Experiments revealed 8 user groups (covering 3 types, such as Drama), and the cumulative regret of this algorithm was significantly lower than that of the ordinary UCB, with the regret growth slowing down in the middle and high rounds. The personalized α adapts to the exploration needs of different groups. This research provides a reusable framework for heterogeneous scenario recommendations and has significant practical value.
