Online Algorithm for Unsupervised Sequential Selection with Contextual Information
This addresses a challenging unsupervised learning problem in sequential decision-making with contextual information, but it is incremental as it builds on existing bandit frameworks with a new variant.
The paper tackles the Contextual Unsupervised Sequential Selection problem, a variant of stochastic contextual bandits where loss cannot be inferred from feedback, by proposing an algorithm under the Contextual Weak Dominance property and showing it achieves sub-linear regret, with validation on synthetic and real datasets.
In this paper, we study Contextual Unsupervised Sequential Selection (USS), a new variant of the stochastic contextual bandits problem where the loss of an arm cannot be inferred from the observed feedback. In our setup, arms are associated with fixed costs and are ordered, forming a cascade. In each round, a context is presented, and the learner selects the arms sequentially till some depth. The total cost incurred by stopping at an arm is the sum of fixed costs of arms selected and the stochastic loss associated with the arm. The learner's goal is to learn a decision rule that maps contexts to arms with the goal of minimizing the total expected loss. The problem is challenging as we are faced with an unsupervised setting as the total loss cannot be estimated. Clearly, learning is feasible only if the optimal arm can be inferred (explicitly or implicitly) from the problem structure. We observe that learning is still possible when the problem instance satisfies the so-called 'Contextual Weak Dominance' (CWD) property. Under CWD, we propose an algorithm for the contextual USS problem and demonstrate that it has sub-linear regret. Experiments on synthetic and real datasets validate our algorithm.