Introduction
Predicting student success in college is a complex task, far exceeding the scope of single-course performance models. Modern "open curricula" allow for diverse elective paths, creating a high-dimensional and sparse state space. Furthermore, a student's academic trajectory is dynamic; learning habits and goals evolve over four years. Traditional "snapshot" models fail to capture this temporal evolution and often become obsolete as academic environments change. This research introduces a progressive prediction framework designed to forecast on-time graduation by continuously updating its model as students complete each academic quarter.
Methodology: Ensemble-based Progressive Prediction (EPP)
The core of the proposed system is the Ensemble-based Progressive Prediction (EPP) algorithm. It utilizes a bilayered architecture to manage the temporal nature of student data without excessive computational overhead.
- Base Layer: Employs "local" predictors (e.g., K-Nearest Neighbors, SVMs) to analyze a student's academic snapshot for a specific quarter.
- Ensemble Layer: Uses an Exponentially Weighted Average Forecaster (EWAF) to synthesize the current base prediction with the ensemble prediction from the previous quarter. This recursive structure allows the model to "remember" the student's trajectory.
To handle course variety, the researchers applied Domain Knowledge Clustering, grouping 453 courses into 10 clusters (e.g., Math, Physics). LASSO-based feature selection then identified the most significant clusters for prediction at various degree stages.
Data and Features
The study analyzed a real-world dataset from the UCLA Mechanical and Aerospace Engineering Department, encompassing 367 anonymized students. Features included high school GPA, SAT scores, course credits, and grades across 453 distinct courses. The target variable was on-time graduation (within 12 quarters), which had a 58% rate in the dataset.
Key Insights
Analysis revealed several critical factors in student success:
- Trajectory Matters: The ensemble approach captures the evolution of learning styles, performing as well as the best "expert" in hindsight.
- SAT vs. HS GPA: Within this engineering cohort, SAT scores were found to be more significant static predictors than high school GPA.
- Quality of Learning: Course grades were more predictive than the total number of credits earned, suggesting that mastery is more important than volume.
- Temporal Feature Importance: The relevance of specific features shifts over time; for instance, "Senior Design" grades become critical in later quarters.
Results and Performance
The EPP algorithm was benchmarked against GPA-only, Credit-only, and "Base" (snapshot-only) models.
The EPP model consistently outperformed all benchmarks. As shown in the results, accuracy improves progressively as more data is collected. By the 12th quarter, the model achieved near-perfect classification for high-SAT students. Crucially, the model provides an "early warning" advantage as early as the 4th to 6th quarter, offering a significant window for academic intervention.
Conclusion and Future Work
While effective in an engineering context, future work should explore the model's performance in more flexible liberal arts programs. Additionally, integrating non-academic factors like socio-economic status could further refine predictions. Ultimately, this data-driven approach is intended to augment human pedagogical expertise, helping advisors identify and support at-risk students earlier in their academic careers.
Summary updated