The impact of dataset quality on the performance of data-driven approaches for human activity recognition

This paper discusses the impact of data quality on activity classification using data-driven approaches. Data was collected by 141 undergraduate students at Ulster University using a triaxial accelerometer. A clearly defined data collection protocol was provided for the participants as data collection occurred in an unsupervised setting. Results produced by four common classifiers highlight the effects of noisy data by comparing the classification performances of raw and subsequently cleaned data. Results also highlight the importance of following a data collection protocol attentively to improve overall classification performance. The Naive Bayes classifier improved most significantly with a 12.967% increase in performance.