Lexi-LowGLM: Efficient Online Algorithm for Multi-Objective Matrix Bandits

A recent paper published on arXiv (2608.04324) presents Lexi-LowGLM, a novel online algorithm designed for generalized low-rank matrix bandits that accommodate multiple prioritized objectives. This research explores a situation where, in each round, a learner picks a matrix-valued arm and receives a vector-valued reward, with components reflecting various objectives of differing priority levels. Each objective is represented by its own generalized low-rank matrix model, and the learner assesses arms based on a lexicographic order, emphasizing higher-priority objectives first. Lexi-LowGLM begins by estimating low-rank subspaces specific to each objective, followed by lexicographic learning within these reduced feature spaces. Unlike traditional single-objective algorithms that utilize all past observations to solve a batch generalized linear estimator, Lexi-LowGLM employs an online Newton step to update each objective-specific estimator, thereby decreasing the complexity of updates. This work, contributed by researchers and available on arXiv, advances the domain of online learning and bandit algorithms, with implications for recommendation systems and sequential decision-making involving multiple objectives.