PAT: A power-aware decision tree algorithm for mobile activity recognition

Mobile context recognition attempts to infer the context of a mobile phone user. Machine learning algorithms exhibit high classification accuracy in these applications. Most approaches are very power-inefficient because they record data from all sensors at all time. Intelligently cycling sensors could greatly improve the power efficiency of context recognition services. We propose a decision tree-based machine learning algorithm which optimizes not only on classification accuracy, but also on data retrieval costs based on power efficiency. We show that in a simple physical activity recognition application, the use of this new algorithm results in a significant decrease in power consumption while maintaining a high classification accuracy.