Stability-Aware Counterfactual Recourse For Unsupervised Educational Clustering

Unsupervised clustering is widely used in Learning Analytics to identify latent student behavioral profiles, yet these analyses often remain descriptive and rarely support actionable decision-making. Counterfactual recourse provides a means to identify minimal behavioral adjustments that enable transitions between clusters, though stable clustering partitions do not necessarily imply stable recommended actions. This study introduces a stability-aware evaluation protocol for counterfactual recourse in unsupervised educational clustering. The framework combines K-means behavioral profiling, bootstrap-based cluster stability assessment using the Adjusted Rand Index (ARI), and assignment-consistent counterfactual generation under feasibility and typicity constraints. Experiments are conducted on two OULAD presentations (2013B and 2014B), selected according to study-design criteria defined prior to the main analysis, and on a pruned version of the KDD Cup 2015 dataset used as an external validation set. Two aligned behavioral transitions are considered: Explorers to Participants and Participants to Immersives. Evaluation considers validity, coverage, cost (median/p90), sparsity, partition stability, validity retention, and action stability measured through Jaccard similarity under bootstrap retraining. Across settings, recourse remains highly feasible overall, but robustness is heterogeneous across datasets and transition types. In particular, stable partitions do not uniformly translate into stable action sets or cost-robust recommendations. These results reinforce the paradox of stability and show that stability-of-recourse should complement traditional clustering robustness metrics in actionable educational clustering. In practice, this enables the identification of behavioral recommendations that are not only feasible but also reliable under model retraining.