Fairness-aware Machine Learning in Educational Data Mining

Abstract

Fairness is an essential requirement of every educational system, which is reflected in a variety of educational activities. With the extensive use of Artificial Intelligence (AI) and Machine Learning (ML) techniques in education, researchers and educators can analyze educational (big) data and propose new (technical) methods in order to support teachers, students, or administrators of (online) learning systems in the organization of teaching and learning. Educational data mining (EDM) is the result of the application and development of data mining (DM), and ML techniques to deal with educational problems, such as student performance prediction and student grouping. However, ML-based decisions in education can be based on protected attributes, such as race or gender, leading to discrimination of individual students or subgroups of students. Therefore, ensuring fairness in ML models also contributes to equity in educational systems. On the other hand, bias can also appear in the data obtained from learning environments. Hence, bias-aware exploratory educational data analysis is important to support unbiased decision-making in EDM. In this thesis, we address the aforementioned issues and propose methods that mitigate discriminatory outcomes of ML algorithms in EDM tasks. Specifically, we make the following contributions: We perform bias-aware exploratory analysis of educational datasets using Bayesian networks to identify the relationships among attributes in order to understand bias in the datasets. We focus the exploratory data analysis on features having a direct or indirect relationship with the protected attributes w.r.t. prediction outcomes. We perform a comprehensive evaluation of the sufficiency of various group fairness measures in predictive models for student performance prediction problems. A variety of experiments on various educational datasets with different fairness measures are performed to provide users with a broad view of unfairness from diverse aspects. We deal with the student grouping problem in collaborative learning. We introduce the fair-capacitated clustering problem that takes into account cluster fairness and cluster cardinalities. We propose two approaches, namely hierarchical clustering and partitioning-based clustering, to obtain fair-capacitated clustering. We introduce the multi-fair capacitated (MFC) students-topics grouping problem that satisfies students' preferences while ensuring balanced group cardinalities and maximizing the diversity of members regarding the protected attribute. We propose three approaches: a greedy heuristic approach, a knapsack-based approach using vanilla maximal 0-1 knapsack formulation, and an MFC knapsack approach based on group fairness knapsack formulation. In short, the findings described in this thesis demonstrate the importance of fairness-aware ML in educational settings. We show that bias-aware data analysis, fairness measures, and fairness-aware ML models are essential aspects to ensure fairness in EDM and the educational environment.