Predicting App Usage from Mobile Internet Traffic Data Based on Node and Path Similarity

Abstract

Due to the large number of mobile applications (Apps) installed on the smartphones, it becomes time-consuming for the user to find the Apps he or she want to use, therefore, it is important to predict App usage based on understanding of mobile user's behavior. In this paper, we use mobile Internet traffic data to construct User-App bipartite network, and transform the App usage prediction task to a link prediction problem. To better mine the relations between users and Apps, we propose a new node based and two path based similarities that are bidirectional conditional probability (BCP) similarity based on preference property of nodes, local shortest path (LSP) similarity based on closeness relationship between nodes, and random walk with resource redistribution (RWRR) similarity based on the node degree. We propose an App usage prediction framework BLR-AUP to predict which App the user will use. Experiment results showed that our model outperformed other traditional link prediction models, with high F1 and AUC value of 91.14% and 96.43%.