Clustering and Naïve Bayesian Approaches for Situation-Aware Recommendation on Mobile Devices

In this paper, we target the problem of the situation-aware application (task) recommendation on mobile devices. To tackle this problem, we develop both supervised and unsupervised approaches. We use Naive Bayesian as a supervised approach, and co-clustering and vector quantization (VQ) as unsupervised approaches. We evaluate the performance of the proposed approaches with both synthetic and actual user log data that we have collected for six months. Our initial experiment shows that the co-clustering-based approach results in comparable purity performance with much less computation time than VQ. Therefore, the co-clustering approach can be practical for high dimensional data. Furthermore, we characterize the recommendation performance of the proposed approaches in terms of the receiver-operating-characteristics (ROC). One interesting observation is that the unsupervised approaches perform well with a single identical threshold over all applications, while the supervised approach does better with a different threshold for each application.