Sequence-Graph Fusion Neural Network for User Mobile App Behavior Prediction

Yizhuo Wang, Renhe Jiang, Hangchen Liu, Du Yin, Xuan Song

Abstract: In recent years, mobile applications (apps) on smartphones have shown explosive growth. Massive and diversified apps greatly affect user experience. As a result, user mobile app behavior prediction has become increasingly important. Existed algorithms based on deep learning mainly conduct sequence modeling on the app usage historical records, which are insufficient in capturing the similarity between users and apps, and ignore the semantic associations in app usage. Although some works have tried to model from the perspective of graph structure recently, the two types of modeling methods have not been combined, and whether they are complementary has not been explored. Therefore, we propose an SGFNN model based on sequence combined graph modeling, which is already publicly available as the GitHub repository https://github.com/ZAY113/SGFNN . Sequence Block, BipGraph Block, and HyperGraph Block are used to capture the user mobile app behavior short-term pattern, the similarity between users and apps, and the semantic relations of hyperedge “user-time-location-app”, respectively. Two real-world datasets are selected in our experiments. When the app sequence length is 4, the prediction accuracy of Top1, Top5, and Top10 reaches 36.08%, 68.39%, 79.02% and 51.55%, 87.57%, 95.62%, respectively. The experimental results show that the two modeling methods can be combined to improve prediction accuracy, and the information extracted from them is complementary.