A gradient descent matrix factorization for microtask crowdsourcing

Conventional crowdsourcing algorithms assign all tasks to all workers and collect their final answers. However, a major issue with these baseline approaches is their inability to identify and eliminate answers from low-quality or non-genuine workers. To tackle this challenge, novel microtasking algorithms have been proposed where workers' expertise is estimated to allocate tasks accordingly. These algorithms rely on a worker-task matrix, but a common problem is the potential sparsity of this matrix due to not all workers answering allocated tasks and a high ratio of tasks to workers. To address this issue, this paper introduces a novel optimization-based matrix factorization approach using Gradient Descent known as GRADi. GRADi aims to predict missing answers from different workers while factorizing the worker-task matrix. During this process, GRADi incorporates worker similarity information to enhance factorization accuracy. The effectiveness of GRADi is evaluated using datasets from Amazon Mechanical Turk and similar platforms, assessing accuracy and Root Mean Square Error. Comparative analyses against other matrix factorization algorithms, including both optimization and non-optimization techniques, as well as existing microtasking algorithms, demonstrate that GRADi consistently outperforms these methods. It shows promising results in improving microtasking outcomes by better predicting worker responses and handling the inherent challenges of sparse worker-task matrices.