Explicative Human Activity Recognition using Adaptive Association Rule-Based Classification

Computational social sensing is enabled by the Internet of Things at large scale. Using sensors, e.g., implemented in mobile and wearable devices, human behavior and activities can then be investigated, e.g., using according models and patterns. However, the obtained models are often not explicative, i.e., interpretable, transparent, and explanation-aware, which makes assessment and validation difficult for humans. This paper proposes a novel explicative classification approach featuring interpretable and explainable models. For this purpose, we embed a framework for building rule-based classifiers using class association rules. For evaluation, we apply two real-world datasets: One collected in the domain of personalized health using wearable sensors (accelerometers), the second one utilizing smartphone sensors for activity recognition. Our results indicate, that the proposed approach outperforms the baselines clearly, concerning both accuracy and complexity of the resulting predictive models.