Introduction
In Massive Open Online Courses (MOOCs), the sheer scale of enrollment—often reaching thousands of students—precludes individualized instructor attention. While clustering students into behavioral groups is a standard approach to managing this scale, it faces two significant challenges: the curse of dimensionality and homogenization bias. Modern educational platforms track a vast array of clickstream data, creating high-dimensional datasets where traditional clustering becomes noisy. Furthermore, most feature selection methods are "global," assuming that the same behaviors drive success for every student. This ignores the diversity of learning styles inherent in a global student body.
To address these issues, we present Interpret3C (Interpretable Conditional Computation Clustering), a novel pipeline that performs individualized feature selection to create highly interpretable and actionable student clusters.
The Interpret3C Framework
Interpret3C operates through a three-stage process designed to respect individual differences while remaining scalable for large datasets.
- Individualized Feature Selection: Rather than manual selection, the system employs an interpretable neural network architecture. A "discriminator layer" utilizes a Gumbel-SoftMax gating mechanism to act as a mask, selecting a specific subset of features for each student based on their unique behavioral patterns.
- Supervised Proxy Training: To ensure the selected features are educationally relevant, the network is trained on a proxy task: predicting academic success (pass/fail). This grounds the feature selection in tangible learning outcomes.
- Spectral Clustering: After generating individual feature masks, the system performs Spectral Clustering. The distance between students is calculated using only their "activated" (important) features, ensuring clusters are formed based on the most relevant signals for those specific individuals.
Experimental Evaluation
The model was evaluated using data from a Master’s level Digital Signal Processing MOOC, involving 5,611 active students over a 10-week period. The study focused on the first four weeks of clickstream data—including video interactions, quiz performance, and session regularity—to simulate an early-intervention environment.
Key Results and Interpretability
Interpret3C identified six distinct behavioral clusters (A through F). A critical finding was the "Power of Eight": out of 45 original features, only eight were found to be significant across the population. Five of these were nearly universal, while three varied significantly between clusters.
One of the most striking insights involved "Content Alignment" (adherence to the course schedule). The model flagged this as a critical feature primarily for students who were falling behind. For high-achievers, the model often ignored this feature, as their success was driven by other factors like intense video engagement.
By visualizing these feature masks, educators can see exactly why a student was placed in a specific cluster. For example, while Clusters D and E both had low passing rates (~16%), Cluster D showed high quiz engagement while Cluster E was generally disengaged. A global model might have grouped them together, but Interpret3C reveals they require different support strategies.
Conclusion and Future Work
Interpret3C demonstrates that individualized feature selection can significantly enhance the interpretability of student clusters in MOOCs. By moving away from "one-size-fits-all" feature importance, the model uncovers nuanced learning patterns that are otherwise lost in the noise. Future work will explore the generalizability of these "Universal 5" features across different academic domains and investigate how changing the proxy task (e.g., from pass/fail to retention) alters the selected feature sets.
Summary updated