[EDM 2018] Enhancing Student Performance Prediction: An Ensemble-based Decision Tree Approach

TL;DR

Predicting student success is no longer just about grades; it’s about uncovering the hidden variables in the learning journey. This research introduces a methodology that marries the interpretability of Decision Trees (RandomTree and REPTree) with the robust predictive power of the Rotation Forest ensemble algorithm. The result? A significant boost in accuracy (up to 93.02%) and the generation of highly detailed, actionable educational rules.

Background: The Shift to Educational Data Mining (EDM)

As educational platforms move online, the "Explosion of Information" has left educators with mountains of raw data. The challenge isn't just storing it, but converting it into a compass for student success. This paper positions itself as a solution for Educational Data Mining (EDM), focusing on the User Knowledge Modeling dataset to classify students' expertise levels—from "Very Low" to "High."

The Core Motivation: Interpretability vs. Accuracy

Standard Decision Trees like REPTree are beloved in education because they provide "Rules"—logic paths like IF Study_Time is High THEN Grade is High. However, these simple models often miss subtle interactions between variables like exam performance for goal objects (PEG) and repetition numbers (SCG). The authors sought to fix this by using Rotation Forest, an ensemble method that doesn't just average results but actually "rotates" the feature space to help trees see the data from different, more informative angles.

Methodology: The Rotation Forest Synergy

The methodology follows a two-stage pipeline:

1. Base Classification: Establishing baselines using RandomTree and REPTree.
2. Ensemble Optimization: Wrapping these trees in a Rotation Forest.

Rotation Forest is unique because it uses Principal Component Analysis (PCA) to rotate the feature axes for each subset of trees. This forces each tree in the ensemble to focus on different linear combinations of student attributes, leading to a "forest" that is far more knowledgeable than any single "tree."

Table: Performance of the baseline RandomTree across various metrics.

Experimental Insights & Results

The researchers tested their approach on data from Gazi University, Turkey. The findings were conclusive:

• Accuracy Leap: Simple RandomTree started at 88.37%. By applying Rotation Forest (RF-RandomTree), accuracy jumped to 92.25%.
• The SOTA Winner: RF-REPTree emerged as the strongest model with 93.02% accuracy.
• Complexity is Content: Interestingly, the size of the trees increased dramatically (from 15 nodes in a simple REPTree to 190 in the ensemble version). In EDM, a bigger tree often means more refined, specific rules for different student personas.

Table: Comparison of accuracy, tree size, and running time across simple and ensemble methods.

Beyond Metrics: Understandable Rules

One of the most valuable outputs of this research is the generation of logical rules. For example, the ensemble models provided complex triggers involving multiple variables like LPR (Exam performance for related objects) and PEG. These rules allow teachers to identify exactly why a student is classified at a certain level, enabling targeted intervention.

Conclusion & Future Outlook

This paper demonstrates that ensemble techniques are not just "black boxes." When applied to decision trees, they enhance the granularity of the information extracted.

Limitations: While the accuracy is high, the model's complexity (600+ nodes) might make manual rule-checking difficult for human educators. Future Work: The next step in this field involves applying these ensemble methods to real-time e-learning environments where the model can adjust teaching strategies on the fly based on student behavior data.

Takeaway for Practitioners: If your current classification model is underperforming, don't just switch models—try wrapping your existing decision trees in a Rotation Forest. The accuracy gain is often worth the slight increase in computational time.