Payment behavior prediction on shared parking lots with TR-GCN

Shared parking lots are new types of sharing economy and generate a large social impact in our daily lives. Post-use payment is a hallmark method in the shared parking lots: it reflects trust in users and brings convenience to everyone. Accordingly, payment behavior prediction via data science technology becomes extremely important. We cooperate with a real intelligent parking platform, ThsParking, which is one of the top smart parking platforms in China, to study payment prediction, and encounter three challenges. First, we need to process a large volume of data generated every day. Second, a variety of parking related data shall be utilized to build the prediction model. Third, we need to consider the temporal characteristics of input data. In response, we propose TR-GCN, a temporal relational graph convolutional network for payment behavior prediction on shared parking lots, and we build a reminder to remind unpaid users. TR-GCN addresses the aforementioned challenges with three modules. 1) We develop an efficient data preprocessing module to extract key information from big data. 2) We build a GCN-based module with user association graphs from three different perspectives to describe the diverse hidden relations among data, including relations between user profile, temporal relations between parking patterns, and spatial relations between different parking lots. 3) We build an LSTM-based module to capture the temporal information from historical events. Experiments based on 50 real parking lots show that our TR-GCN achieves 91.2% accuracy, which is about 7% higher than the state-of-the-art and the reminder service makes more than half of the late-payment users pay, saving 1.9% loss for shared parking lots.