When Labor-Intensive Mobile Crowdsourcing Meets Unobservability: Contextual Bandit Learning with Unobservable Individual Rewards

Mobile crowdsourcing (MCS) has emerged as an effective means of leveraging the power of the crowd for large-scale location-related tasks. However, one key challenge of labor-intensive MCS is assigning labor-intensive tasks to suitable workers, as different workers are better suited for different tasks and contribute differently to the overall goal. Moreover, these relationships are often unknown and time-varying, and decision-makers typically focus on the overall task performance rather than individual worker performance. Previous works have not addressed this problem adequately, as they either relied on individual rewards for decision-making or assumed a known relationship between individual rewards and the overall reward. To address this problem, we propose a new approach that models labor-intensive task assignments as a contextual bandit learning problem with unobservable individual rewards (UIR). Our approach employs an improved UCB-UIR algorithm for the known reward relationship scenario, which yields a sublinear regret bound with UIR. For the unknown reward relationship scenario, we propose a generic Transformer-UIR algorithm to learn the relationship between contextual information and then determine optimal assignments with UIR. We demonstrate the effectiveness of our approach using a realistic MCS application, where our algorithms outperform state-of-the-art baselines significantly with both known and unknown reward relationships.