Time-aware GCN: Representation Learning for Mobile App Usage Time-series Data

Abstract
With the expansion of smartphones, most users are using many apps (applications) on their smartphones. As the way users use their apps would reflect their personality, understanding their app usage is increasingly becoming an interesting problem. In order to understand their app usage, which consists of time-series data, we can use sequential models such as N-gram and long short-term memory (LSTM) for considering sequence characteristics. However, it is still challenging to reduce the impact of internal factors (e.g., their feelings) and external factors (e.g., notifications on their smartphones) on the differences in the order of apps used in a short term. In this paper, we propose a novel method for representation learning of app usage, called Time-aware Graph Convolutional Networks (T-GCN), to address the problem mentioned above. We evaluated the performance of T-GCN with the largescale real-world dataset on the app usage prediction task. The results demonstrate that T-GCN achieves 3.6% higher accuracy than the LSTM model in accuracy@10.