Machine Learning

Adaptive Exam Preparation System using Active Learning

An active-learning system that estimates a student's ability level in as few exam questions as possible. By choosing the most informative question at each step, it reaches the same confident assessment while asking 60% fewer questions than a fixed quiz.

Project Overview

Traditional assessment hands every student the same long, fixed quiz. Adaptive Exam Prep reframes the problem: what is the minimum number of questions needed to confidently estimate a student's ability? Built on the ASSISTments 2009/2010 dataset (~120 MB, ~1,900 students), it treats question selection as an active-learning problem, pick the next question that tells us the most, then re-predict.

The engine pairs Item Response Theory (a 2PL model fit with Expectation-Maximization) with five active-learning query strategies: IRT-Fisher Info, entropy, uncertainty sampling, query-by-committee, and a random baseline, each driving RandomForest, GradientBoosting, and LogisticRegression classifiers retrained at checkpoints k ∈ {2, 4, …, 30}.

Key Features

Active-learning question selection: five query strategies compared (IRT-Fisher information, entropy, uncertainty sampling, query-by-committee, random baseline)
Item Response Theory engine: a 2PL model fit with EM, used to estimate student ability (θ) and drive Fisher-information question selection
60% fewer questions: reaches 80% accuracy at k=12 vs the baseline's k=30, and 86.4% accuracy by k=30 (5-fold CV on 490 held-out students)
Rigorous evaluation: 5-fold cross-validation, an independent external validation set, and four formal success gates the top models all pass
Live Streamlit demo: adaptive vs random run side-by-side on the same student, with a real-time confidence chart marking the 80% threshold