Contextual Thinking for Inference and Prediction of Daily Activities by Mining Smartphone Data

This paper presents a contextual thinking framework for inference and prediction of daily activities by mining smartphone data. A significant activity is a predefined activity to be inferred, for example, waiting for a bus, having a meeting, working in office, taking a break in a coffee shop et al. In this framework, smartphone contextual information consists of local time, user state as well as user significant location. This contextual triad forms a contextual tuple at each time epoch. The activity inference engine is developed using the dynamic Naive Bayes classifier (DNBC) method, which extends a standard hidden Markov model (HMM) to make real-time inference less expensive and to maximize the flexibility of the overall framework for adopting adaptive context tuple elements and significant activity states. As for significant activity prediction, the solution is designed to forecast upcoming user activities based on virtual contextual tuple observation generated by mining historical significant locations and user states. This is achieved in two steps: 1) build virtual contextual observations and maintain a context data file which continuously record the user's contextual information based on the Naive Bayes strategy, but in a real time algorithm; 2) use the DNBC method for activity prediction. The performance of the overall contextual thinking framework was evaluated at the campus of Texas A&M University Corpus Christi. A contextual thinking engine for user significant activity inference and prediction is developed based on the Android platform. Six significant activities were defined and tested by three different testers for a full week at TAMUCC campus using three different Samsung smartphones. Experiment results reveal that the activity inference process can reach an average accuracy level of 78%. Prediction accuracy is relatively poor due to limited data length and low time quantization level.