Energy Consumption of Batch and Online Data Stream Learning Models for Smartphone-based Human Activity Recognition

Online Smartphone-based Human Activity Recognition (SHAR) systems are increasingly used in different kinds of applications such as in smart homes and health monitoring. In order to allow users to train their own personalized models quickly and without privacy concerns, a SHAR model has to be lightweight due to the limited resources of mobile devices. Therefore, time, memory and energy costs have to be analyzed before integrating the learning process in on-device scenarios. In a previous paper, we presented a comparative study for different machine learning (ML) algorithms, regarding training time, accuracy and memory usage, exploiting two different learning approaches: i) CPU-based vs GPU-based batch learning, and ii) Online data stream learning. In this paper, we complete our study by analyzing power consumption for both approaches. Results show that the same previous best models show also the best power values. The impact that has the CPU on the energy cost is also highlighted.