Predicting application usage based on latent contextual information

Predicting application usage is useful for offering personalized services, improving mobile energy consumption, and mobile system resource management optimization. Currently, however, there are many possible applications, and each user has his/her own preferences and usage patterns, which makes the application prediction task very challenging. In this study we use different representation methods to represent mobile users’ contextual information in order to predict application usage. We focus on the spatial information context (i.e., where the applications are used) and represent it with graph embeddings, which capture the locations users have visited based on their movement. We use multimodal embeddings to represent the temporal context, users’ identifiers, and previously used applications. Then, the contextual information's latent representation is used in a deep learning framework composed of a GRU (gated recurrent unit), attention layer, and a softmax layer to provide application usage predictions. We evaluate our method on two real-world datasets comprised of data collected from mobile users’ devices. Our results show that the proposed application usage prediction method outperforms various machine learning models and state-of-the-art solutions. We also found that the spatial information's latent representation derived from graph embeddings outperformed traditional and commonly used representation methods when predicting application usage. Our findings also reveal interesting usage patterns regarding users’ predictability, which can help us better understand users’ behavior.