Disposable Linear Bandits for Online Recommendations

We study the classic stochastic linear bandit problem under restriction that each arm may be selected for limited number of times. This simple constraint, which we call disposability, captures a common occurs in recommendation problems from diverse array applications ranging personalized styling services to dating platforms. show regret this is characterized by previously-unstudied function reward distribution among optimal arms. Algorithmically, our upper bound relies on an optimism-based policy which, while computationally intractable, lends itself approximation via fast alternating heuristic initialized with similarity score. Experiments dominates set benchmarks includes algorithms known without along natural modifications these disposable setting.