Detecting Label Errors in Crowd-Sourced Smartphone Sensor Data

Applications relying on supervised learning algorithms are susceptible to producing false outputs in the presence of label errors, i.e., situations were labels have been corrupted, both deliberately and accidentally. While prior work has focused on detecting and handling label errors for various types of applications, there is a lack of research addressing label errors in smartphone-based crowd-sensing applications, especially when used for action recognition. In this paper, we discuss and address two common types of smartphone-based label errors:mislabeling and multi-action labels. We also compare multiple learning algorithms, including an ensemble of four stratified trained classifiers. The results indicate the importance of the action type for filtering label error. The goal of this work is to provide guidelines for developing effective techniques to discover and remove error labels for action recognition systems.