Enhancing App Usage Prediction Accuracy With GCN-Transformer Model and Meta-Path Context

Abstract

In this paper, we introduce MP-GT, a novel Graph Neural Network model that leverages meta-path-guided optimization within the GCN-Transformer framework to enhance application (App) usage prediction accuracy. Our approach addresses issues such as suspended animation and over-smoothing by extracting both local subgraph structures and global graph structures using a combination of GCN and Transformer method. Furthermore, we enhance the capture of semantic information and App usage patterns by incorporating a meta path-guided objective function. Extensive experiments demonstrate that MP-GT outperforms the widely adopted baseline of semantic-aware representation learning via Graph Convolutional Network (SA-GCN) by 13.33% in terms of Accuracy@1. Additionally, MP-GT surpasses the popular baseline of context-aware App usage prediction with heterogeneous graph embedding (CAP) by 74.02% in the same metric. Moreover, MP-GT reduces training time by 79.47% compared to SA-GCN. These findings validate that our approach not only achieves higher prediction accuracy but also converges faster than the baseline models. Therefore, MP-GT proves to be an effective and superior solution for the App usage prediction task.

Keywords

• Heterogeneous graph embedding
• GCN-transformer
• meta-path optimization
• application usage modeling