The Impact of Data Reduction on Wearable-Based Human Activity Recognition

One crucial step toward improving any pattern recognition model is refining the data (feature extraction) and simplifying it (feature selection) for the classifier. In this paper, we investigate the impact of feature reduction on the performance of HAR. We collected step data from two subjects and answer research questions related to the impact of feature reduction in terms of performance, generalizability and varying classifiers. Our findings indicate feature reduction can reduce the number of features by close to 90%, while only having an impact of 1-2% in model performance. Moreover, we find that feature reduction can impact the generalizability of HAR models. Lastly, we find that feature reduction does not have a major impact on most classifiers examined. Our results are useful for designers of HAR systems to help them optimize their models while ensuring high performance.