Crowd-CapsNet: capsule network based adaptive crowdsourcing task recommendation

Purpose The crowdsourcing software development platforms organize geographically distributed developers to complete various developing tasks, bringing convenience and efficiency to users. However, with the increasing number of both developers and tasks, it becomes more and more challenging to match tasks and suitable developers, especially for imbalanced data. The purpose of this paper is to propose an accurate and diverse recommendation model for crowdsourcing tasks. Design/methodology/approach A revised circle loss function is applied to achieve a certain adaptive ability, which is critical for imbalanced data, it guarantees diversity by maximizing the target label score and leveraging mathematical approximation to automatically balance the weights. Besides, the authors leverage the capsule network to obtain the semantic feature of tasks’ descriptions, modify the dynamic routing mechanism to better learn users’ preferences and improve the recommendation accuracy. Findings The comprehensive experiments conducted on real crowdsourcing platform data demonstrate that the proposed Crowd-CapsNet model can achieve high recommendation accuracy with a certain diversity. It improves around 1% accuracy with only 37% training time of the LSFA approach. Originality/value This paper proposes Crowd-CapsNet, an adaptive crowdsourcing task recommendation model. A relatively general feature pre-processing method describes crowd-sourcing tasks and the modified capsule network further obtains the semantic features to improve the recommendation accuracy and diversity.