Is Querying Users Acceptable for Human Activity Recognition Based on Active Learning?

Labeled data scarcity is one of the major open problems of sensor-based Human Activity Recognition (HAR). To mitigate this issue, several research groups proposed solutions based on Active Learning (AL), where the user is explicitly asked to provide feedback about the performed activity when the classifier's confidence is low. Despite existing methods trigger a limited number of queries that decreases over time, they do not take into account user's context. Indeed, a subject may not wish to be interrupted when receiving the query's notification. Delaying the query may be critical, since it is known that the feedback's quality decreases with the length of the delay. In this work, we aim at answering the following question: is real-time AL-based HAR practical in real-world scenarios? By using a novel evaluation methodology based on the Wizard of Oz approach, we performed a user-based study with 30 subjects performing physical activities and using earphones and a smartwatch as vocal and touch interfaces. Our results evaluate the impact of the AL strategy, the user's context, and the interaction modality on the effectiveness of the AL approach.