Introduction
Identifying students at risk of academic failure is a critical Key Performance Indicator (KPI) for higher education institutions. Traditional methods often rely on manual surveys or Learning Management System (LMS) logs. However, surveys are resource-intensive and prone to participation bias, while LMS data is often siloed within specific courses, failing to provide a holistic view of student behavior. This study explores an alternative, non-intrusive data source: university firewall logs. By capturing all campus internet requests, these logs offer an automated and comprehensive perspective on student activity across their entire academic environment.
Methodology
The researchers implemented an Educational Data Mining (EDM) framework to process approximately 70GB of raw firewall logs daily. The approach utilized a "schema on read" strategy to manage semi-structured data efficiently. Feature engineering focused on two primary domains: Duration (total time spent) and Frequency (number of access hits).
The study involved 294 science students living off-campus. Data was categorized into two datasets:
- Dataset-1 (CAT): 24 broad web categories (e.g., social networking, search engines).
- Dataset-2 (APP): 147 specific application types (e.g., Facebook-base, YouTube-base).
Students were classified as "at-risk" if their GPA fell below 2.4. Five machine learning models were evaluated: Naive Bayes, Logistic Regression, Neural Networks, Decision Trees (J48), and Random Forest.
Behavioral Insights
Analysis revealed that the duration of internet sessions is a significantly stronger predictor of academic success than the frequency of access. Specific application data (APP) also proved more descriptive than broad categories (CAT).
A notable finding was the "Gaming Paradox": successful students spent more time on online games (avg. 6.48 min/session) than at-risk students (avg. 1.86 min/session), suggesting that high-performing students may use gaming for structured stress relief. Conversely, at-risk students generally spent more time online per hour (2.72 minutes vs. 2.28 minutes) during the first week of the semester, potentially indicating early procrastination.
Key Results
The Random Forest algorithm emerged as the most effective model, consistently outperforming others across all datasets with an average accuracy of 0.778.
The highest predictive accuracy (0.779) was achieved using the APP-Duration dataset. Further optimization showed that a weighted combination of 90% Application data and 10% Category data provided the most robust identification of at-risk students.
Conclusion and Future Work
While firewall logs provide a powerful "invisible" data source for predicting student outcomes, the study notes limitations such as the small sample size and the short one-week data collection window. Future research aims to extend the analysis over full academic years and explore Deep Learning models to capture more complex behavioral patterns. The ultimate objective is to integrate these predictive insights into real-time academic intervention systems.
Summary updated