Probabilistic Machine Learning for Simulating Complex Learner Profiles

Data is the foundation of all machine learning applications. In education science, especially for the learner characteristics that drive personalized learning, it is difficult to collect and often uncertain. It is challenging to model, train, evaluate, and analyze the underlying algorithms when developing AI-based systems and having small sample sizes. To address these problems, we present a synthetic data generator utilising probabilistic models. This generator can effectively model and simulate complex learner profiles. To achieve this, we collected extensive data on learning styles, learning strategies, personalities, and preferred learning paths from 593 students over several semesters at a higher education level. Then, Bayesian networks, Hidden Markov Models, and Markov Chains are used to model the relationships between learner profiles. Using the Bayesian information criterion, and cross-validation with log-likelihood scores, we compare various models to select the best fitting one for synthesizing the data. The synthetic data is then evaluated using statistical validation techniques. In addition, we developed a simulation module with the option to simulate learner profiles based on manual user-defined inputs. The data and code used in this work are available as open source 1contributing to open science and developers for customized simulated data. In the future, this data will refine the training, evaluation, analysis, and benchmarking of algorithms for personalized learning.