SeqST-ResNet: A Sequential Spatial Temporal ResNet for Task Prediction in Spatial Crowdsourcing

Abstract

Task appearance prediction has great potential to improve task assignment in spatial crowdsourcing platforms. The main challenge of this prediction problem is to model the spatial dependency among neighboring regions and the temporal dependency at different time scales (e.g., hourly, daily, and weekly). A recent model ST-ResNet predicts traffic flow by capturing the spatial and temporal dependencies in historical data. However, the data fragments are concatenated as one tensor fed to the deep neural networks, rather than learning the temporal dependencies in a sequential manner. We propose a novel deep learning model, called SeqST-ResNet, which well captures the temporal dependencies of historical task appearance in sequences at several time scales. We validate the effectiveness of our model via experiments on a real-world dataset. The experimental results show that our SeqST-ResNet model significantly outperforms ST-ResNet when predicting tasks at hourly intervals and also during weekday and weekends, more importantly, in regions with intensive task requests.