Machine Learning in Education
Automated Grading Through Contrastive Learning: A Gradient Analysis and Feature Ablation Approach
This study introduces a novel automated SQL grading system leveraging contrastive learning and explainable AI (XAI) techniques to provide accurate, consistent, and transparent feedback. By comparing student submissions to correct solutions in a high-dimensional latent space, the model predicts error percentages. Feature ablation and integrated gradients then pinpoint specific problematic tokens, offering actionable insights for students and instructors. This approach significantly enhances grading efficiency and educational value, aligning closely with human grading standards while reducing manual effort.
Key Performance Indicators
Our automated grading system delivers measurable improvements in educational efficiency and student learning outcomes.
0.97
Grading Accuracy (r)
10x
Feedback Speed Improvement
34%
Optimal Augmentation Rate
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Core Methodology
Our system employs unsupervised contrastive learning to map student and correct SQL solutions into a high-dimensional latent space. This allows for natural differentiation of coding approaches and variances, without predefined labels, forming the foundational input layer for our predictive model. The Frobenius norm measures the distance between these representations, quantifying code correctness.
Core Methodology
To provide granular, actionable feedback, we integrate feature ablation and integrated gradients (IGs). These techniques pinpoint specific tokens in student code that most significantly impact grading outcomes. Feature ablation assesses the influence of removing tokens, while IG computes each token's contribution to predictions, clarifying the model's decision-making.
Core Methodology
The distance computed in the latent space directly predicts the percentage of points deducted. By identifying the most impactful tokens via XAI, the system generates clear, annotated feedback highlighting errors. This approach aligns closely with human grading standards, providing transparent insights for student improvement and instructor efficiency.
0.02 MSE
Achieved on Validation Dataset (Fine-Tuning)
Automated Grading Workflow
SQL Code Submission
→
Data Tokenization
→
Contrastive Learning (Latent Space)
→
Distance Calculation (Frobenius Norm)
→
Error Percentage Prediction
→
XAI Feedback Generation
| Feature | Our Contrastive + XAI Approach | Traditional APAS (Static/Dynamic) |
|---|---|---|
| Grading Granularity |
|
|
| Feedback Quality |
|
|
| Consistency |
|
|
| Adaptability |
|
|
| Overfitting Risk |
|
|
Real-World Impact: University of Zagreb
At the Faculty of Electrical Engineering and Computing in Zagreb, our system was deployed in 'Introduction to Databases' courses. Over 10,373 SQL submissions were processed between April 2021 and April 2024. The automated grading saved instructors an average of 22 hours per 1500 SQL solutions, previously spent on manual assessment. Students reported significantly clearer feedback, leading to a 15% improvement in assignment completion rates and reduced resubmissions. This demonstrates the system's capacity to handle large-scale academic assessments efficiently and effectively, delivering both time savings for educators and enhanced learning outcomes for students.
Advanced ROI Calculator
Estimate the potential time and cost savings for your organization by automating programming assignment grading with AI.
Estimated Annual Savings
$10,000
Annual Hours Reclaimed
200
Your AI Implementation Roadmap
A structured approach to integrating automated grading into your educational or enterprise workflow.
Phase 1: Discovery & Strategy
Conduct an in-depth analysis of your existing grading processes and learning objectives. Define clear KPIs for AI-driven assessment integration.
Phase 2: Data Preparation & Model Training
Assist in anonymizing and preparing historical student code and grading data. Customize and train the contrastive learning model to your specific programming languages and curriculum.
Phase 3: Integration & Pilot Deployment
Seamlessly integrate the AI grading engine with your existing Learning Management System (LMS). Conduct pilot programs with selected courses to validate performance and gather feedback.
Phase 4: Full-Scale Rollout & Optimization
Deploy the system across all relevant courses. Continuously monitor performance, refine the model with new data, and provide ongoing support and feature enhancements.
Ready to Transform Your Grading Process?
Book a free consultation with our AI specialists to explore how automated grading can benefit your institution.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat