Dynamic Time Warping-Based Detection of Multi-clicks in Button Press Dynamics Data

Dynamic Time Warping (DTW) is a method generally used to align pairs of time series with different lengths, which is for instance applied in speech recognition. In this study, we use a category learning experiment (CLE) as use-case, in which the participants have to learn a specific target from a pool of predefined categories within a certain amount of time. From a companion system-based point of view, it is important to detect certain anomalies related to affective states, such as surprise or frustration, elicited during the course of learning. In this work, we analyse the button press dynamics (BPD) data from an auditory CLE with the goal of detecting anomalies of the aforementioned type. To this end, we first select a small set of participants, for which we have definite ground truth labels. Subsequently, we apply DTW in combination with hierarchical clustering to separate the anomaly-specific data from the remaining samples. We compare the outcomes to clustering results based on the extraction of intuitive task-specific features. Our results indicate that applying the DTW approach in combination with Single Linkage Clustering in order to detect CLE-related anomalies is preferable to its feature extraction-based alternative, in person-independent scenarios.