Abstract
One of the principal challenges in developing robust machine learning (ML) classification algorithms for human activity recognition (HAR) from real-time smart home sensor data is how to account for variations in (1) the activity sequence length, (2) the contribution each sensor has to a specific activity, and (3) the amount of activity class imbalance. Such changes might generate observations that do not conform to expected patterns potentially reducing the efficacy of classification models for HAR over time. In this paper we address the second issue through an investigation of generated ML models for HAR. We both examine the importance of features from smart home data sets to enable accurate learning of activities for a ML algorithm and determine what effect these important features have on each specific type of activity. We achieve these objectives through adoption of embedded feature selection methods to reduce the number of features to generate accurate ML models for HAR. Shapley values extracted from these trained models show the degree of impact these extracted features have on classifying an activity. High average area under the roc curve scores were obtained even on the reduced feature sets. Outcomes of our work not only revealed differences in features selected and their importance depending on the ML model generated but also raises questions around how many sensors are required for robust smart home HAR potentially lowering the cost of smart home systems.