Comparison of context-aware predictive modeling approaches: Semantic place in inferring mobile user behavior

The aim of this paper is to empirically examine how to best incorporate such contextual data, such as location or the semantic place of mobile users, into mobile user behavior models. Acquiring such data has become technically easier than ever. The proper utilization of these data leads, hypothetically, to better understanding of mobile user behavior and, consequently, to enhanced mobile services.

The paper systematically compares, under multiple experimental settings, the predictive performances of models built with three different approaches (pre-filtering, contextual modeling and post-filtering) used for incorporating contextual data into user behavior models. The comparisons focus on by which approach additional semantic place information can be best utilized for making the most accurate inferences on mobile user behavior. Real-life smartphone usage data are utilized in the analysis.

The results demonstrate that none of the considered approaches uniformly dominate the others across all experimental settings. However, they show circumstance specific differences that need to be aligned with practical use cases for the best performance.

Identifying the most suitable approaches for utilizing the semantic place (and other contextual) data is an important practical problem for electronic service providers, whose offerings are increasingly moving to the mobile domain and thus need to respond to the demands of mobility.

The paper constitutes an initial step toward understanding and systematically evaluating different approaches for incorporating semantic place data into modeling mobile user behavior. Practitioners in the mobile service domain can apply the initial results and academics build upon them with more diverse experimental settings.