Does sequence mining uncover generalizable behavioral patterns? A methodological case study

Sequence mining techniques provide powerful tools to uncover patterns from time-stamped action sequences. The information extracted with a chosen method, however, may not necessarily be meaningful or generalize beyond the specific data and task. In this case study, we examined the generalizability of behavioral patterns identified across four tasks from the PISA 2025 Learning in a Digital World domain. We compared two subgroup discovery approaches: clustering based on atomistic features (e.g., sequence length, time on task) and sequence clustering using pairwise similarities. Both approaches identified consistent behavioral subgroups across tasks, revealed similar transition patterns, exhibited comparable cluster-covariate relationships with task scores, prior knowledge, and effort, and showed strong agreement with each other. These findings suggest that both approaches captured similar and generalizable differences in how examinees approached the tasks. From our results, we derive implications for mining action sequence data, strongly advocating for incorporating theoretical considerations even in exploratory analyses.