COM-MABs: From Users' Feedback to Recommendation

Abstract

Recently, the COMbinatorial Multi-Armed Bandits (COM-MAB) problem has arisen as an active research field. In systems interacting with humans, those reinforcement learning approaches use a feedback strategy as their reward function. On the study of those strategies, this paper present three contributions: 1) We model a feedback strategy as a three-step process, where each step influences the performances of an agent ; 2) Based on this model, we propose a novel Reward Computing process, BUSBC, which significantly increases the global accuracy reached by optimistic COM-MAB algorithms -- up to 16.2\% -- ; 3) We conduct an empirical analysis of our approach and several feedback strategies from the literature on three real-world application datasets, confirming our propositions.