Title: Prediction of exam question difficulty and experimental applications based on multi-class machine learning algorithms

Authors: Liu Mingyang; Guo Meng

Addresses: College of Arts and Information Engineering, Dalian Polytechnic University, Dalian, 116499, China ' College of Arts and Information Engineering, Dalian Polytechnic University, Dalian, 116499, China

Abstract: Traditional experience-based methods for estimating exam question difficulty are subjective, slow and often inconsistent. This study investigates a data-driven alternative: applying multiclass machine learning (ML) algorithms to predict question difficulty from the perspective of test designers. Using a large-scale question bank, we compared several ML models and feature configurations and found that LbfgsMaxMulti achieved the best overall performance. The model predicted the difficulty of single-choice (SC), multiple-choice (MC), fill-in-the-blank (FB) and true/false (TF) items with > 95% accuracy. We also evaluated three exam paper generation strategies; LbfgsMaxMulti not only excelled at item level predictions but also produced better balanced exam papers than manual or random assembly, thereby enhancing both fairness and scientific rigour.

Keywords: exam difficulty prediction; machine learning; ML; educational assessment; exam paper generation.

DOI: 10.1504/IJADS.2026.155064

International Journal of Applied Decision Sciences, 2026 Vol.19 No.4, pp.386 - 411

Received: 02 Apr 2025
Accepted: 03 Jun 2025
Published online: 25 Jul 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article