BERT-Based Semantic-Aware Heterogeneous Graph Embedding Method for Enhancing App Usage Prediction Accuracy

Abstract

With the widespread adoption of smartphones and mobile Internet, understanding user behavior and improving user experience are critical. This article introduces semantic-aware (SA)-BERT, a novel model that integrates spatio-temporal and semantic information to represent App usage effectively. Leveraging BERT, SA-BERT captures rich contextual information. By introducing a specific objective function to represent the cooccurrence of App-time-location paths, SA-BERT can effectively model complex App usage structures. Based on this method, we adopt the learned embedding vectors in App usage prediction tasks. We evaluate the performance of SA-BERT using a large-scale real-world dataset. As demonstrated in the numerous experimental results, our model outperformed other strategies evidently. In terms of the prediction accuracy, we achieve a performance gain of 34.9% compared with widely used the SA representation learning via graph convolutional network (SA-GCN), and 134.4% than the context-aware App usage prediction with heterogeneous graph embedding. In addition, we reduced 79.27% training time compared with SA-GCN.