Sequential Pattern Mining Suggests Wellbeing Supportive Behaviors

Abstract

Amidst the headlines about the attention economy and the possible impacts of screen time, research investigating the complex relationship between digital technology usage and wellbeing has gained urgency. Researchers generally use a combination of surveys and automatic tracking tools to gather time and frequency of technology use. However, the focus of data analysis has been on measuring duration and frequency of usage rather than exploring behavioral patterns, possibly better indicators of mood states or stress levels. We propose a methodology for detecting behavioural patterns from digital footprints using a sequence pattern mining algorithm, and using these as features for predicting mood. Results show that our method can be used to analyze the relationship between digital usage and mood, and predict the latter with an accuracy of 80%, significantly above the baseline (71.1%). This method provides another angle to investigate digital technology usage in wellbeing-related research.INDEX TERMS Screen time, app usage, digital footprints, affective computing, mood detection.

Doing and Feeling: Relationships Between Moods, Productivity and Task-Switching

“…This detection system has an accuracy of 70% and kappa score around 0.45. Further, our previous studies [12], [13] showed that digital footprints are also useful to predict mood. In this study, we attempted to explore if the mood prediction accuracy will be improved by adding lifestyle or physical context information.…”

“…In this study, the classifier is binary (positive and negative) because it will be the most useful to users: maximum possible accuracy but meaningful labels (similar to our previous studies [12], [13]). If we use more classes of mood, the mood detection system would have lower accuracy.…”

“…Since RQ2 and RQ4 need mood selfreport data, we calculate both task-switching and productivity 60-minute prior to the self-report timestamps. Moods can fluctuate several times in a day [13] so shorter window size is more appropriate for this analysis. Lastly, RQ3 focuses more on lifestyle factors, therefore the last report in a day is the most accurate since it considers the whole day physical activity, social interaction, workload, etc.…”

“…Similar to the previous study [12], we used the dataset where all mood self-reports preceded by RescueTime data with at least 90% window size. Mood detection with duration-based features require a nearcomplete digital footprints data to perform better [13]. Table IX shows the performance metric of the mood detection model by excluding and including lifestyle factor as features.…”

“…Though these productivity tools are available to the public (e.g., RescueTime and ManicTime), based on our knowledge, only a few studies have utilised them [11]. In our previous studies, we have attempted to use their data to detect mood [12], [13]. We argue that these digital tool usage records contain rich information useful as machine learning features.…”