Most works on Multi-Armed Bandits focus the evaluations of their methods on a global accuracy performance metric. In the case of Contextual Multi-Armed Bandit (CMAB), the existing algorithms claim to eventually provide full personalization. This suggests that their global accuracy metric should reflect each user's individual accuracy. In order to verify this, we consider a novel approach of CMAB assessment focused on the evaluation of individual accuracy and compare it to global accuracy. Based on the results of this comparison highlighting some users far from the average global accuracy we propose a heuristic, Sliding Window LinUCB (SW-LinUCB), aiming at decreasing the amount of very unsatisfied users. SW-LinUCB is an adaptation of the original LinUCB CMAB algorithm combined with a diversification mechanism. It acts as the original LinUCB but penalizes arms which are pulled too frequently in order to play them evenly or fairly among optimum. Our approach is motivated by the requirements of recommender systems, which must reach a good global accuracy and should equally distribute it among individuals. We experiment and discuss the benefits and losses of the proposed method on several real-world datasets.
Paper
Full text
Global versus individual accuracy in contextual multi-armed bandit
Semantic Scholar · Computer Science · 2019
Abstract
Most works on Multi-Armed Bandits focus the evaluations of their methods on a global accuracy performance metric. In the case of Contextual Multi-Armed Bandit (CMAB), the existing algorithms claim to eventually provide full personalization. This suggests that their global accuracy metric should reflect each user's individual accuracy. In order to verify this, we consider a novel approach of CMAB assessment focused on the evaluation of individual accuracy and compare it to global accuracy. Based on the results of this comparison highlighting some users far from the average global accuracy we propose a heuristic, Sliding Window LinUCB (SW-LinUCB), aiming at decreasing the amount of very unsatisfied users. SW-LinUCB is an adaptation of the original LinUCB CMAB algorithm combined with a diversification mechanism. It acts as the original LinUCB but penalizes arms which are pulled too frequently in order to play them evenly or fairly among optimum. Our approach is motivated by the requirements of recommender systems, which must reach a good global accuracy and should equally distribute it among individuals. We experiment and discuss the benefits and losses of the proposed method on several real-world datasets.