Early Prediction of Student Learning Difficulty Using Behavioral Patterns

Problem Statement

Early identification of learning difficulty is crucial in educational platforms.
Instead of predicting outcomes after a student fails, this project focuses on predicting whether a student is likely to struggle with a topic based on prior learning behavior.

The core idea is that learning difficulty depends not only on topic-level performance, but also on overall learning patterns of the student.

---

Dataset

Since no public dataset exactly fits this problem, a synthetic dataset was created to allow controlled experimentation and clear interpretation.

• 50 students
• 10 topics
• 500 total records (student–topic interactions)

Each row represents how a student performed on a particular topic.

Features

• avg_prev_score – average performance before the current topic
• time_ratio – actual time spent divided by expected time
• attempts – number of attempts made
• topic_order – sequence of the topic

---

Approach Overview

This project combines unsupervised learning and supervised learning to build a realistic prediction pipeline.

High-level steps:

1. Capture student learning behavior using clustering
2. Integrate behavioral context into topic-level data
3. Predict learning difficulty using a supervised model
4. Validate improvements using a baseline comparison
5. Interpret model behavior

---

1. Student-Level Behavioral Clustering

Learning behavior is a student-level property, not a topic-level one.

Steps:

• Aggregate topic-level data to create student-level behavior profiles
• Use the following features:
  • Mean previous score
  • Mean time ratio
  • Mean number of attempts
• Apply K-Means clustering
• Perform feature scaling since K-Means is distance-based

The resulting clusters represent different learning styles such as fast learners, average learners, and struggling learners.
At this stage, clusters provide context, not predictions.

---

2. Integrating Behavioral Context

Clustering is performed at the student level, while predictions are made at the topic level.

To combine both:

• Cluster labels are merged back into the topic-level dataset using a left join on student_id

This ensures each topic interaction contains both performance data and student behavioral context.

---

3. Defining Learning Difficulty

A realistic and interpretable rule is used to define learning difficulty.

A student is considered to be struggling on a topic if:

• avg_prev_score < 65
• AND
• time_ratio > 1.2

This creates a binary target variable:

• 1 → struggling
• 0 → not struggling

---

4. Difficulty Prediction (Supervised Learning)

To predict learning difficulty, Logistic Regression is used because:

• The target variable is binary
• The model is interpretable
• It aligns with foundational machine learning principles

Training process:

• Train–test split to evaluate generalization
• Feature scaling applied only to input features (to avoid data leakage)
• Model trained using:
  • Topic-level performance indicators
  • Effort-related features
  • Student learning behavior cluster

---

5. Baseline Comparison

To validate whether clustering added value:

• A baseline model was trained without the cluster feature
• Its performance was compared with the cluster-aware model

Observations

• Recall for struggling students remained the same
• Precision improved when behavioral clustering was included

This indicates that learning behavior context reduced false positives without missing at-risk students.

---

6. Model Interpretation

Logistic regression coefficients were analyzed to understand model behavior.

Key insights:

• Higher time spent strongly increases difficulty risk
• Strong prior performance reduces difficulty risk
• Learning behavior clusters add meaningful contextual information

These observations confirm that the model learned reasonable and explainable patterns.

---

Key Takeaways

• Combining unsupervised and supervised learning improves prediction quality
• Behavioral context helps reduce noisy predictions
• Baseline comparison is essential for validating ideas
• Model interpretation is as important as evaluation metrics

---

Future Improvements

• Probability-based early warning thresholds
• Better handling of class imbalance
• API deployment using FastAPI
• Evaluation on real-world educational datasets