Project information
Algorithmic Biases in Machine Learning Models in Education (GAČR PIF - Švábenský)

In machine learning (ML), algorithmic bias can arise at any stage of the modeling pipeline, substantially influencing the resulting models. Previous research observed instances of bias in data and commonly used techniques. However, past work did not describe how the bias changes when parts of the ML pipeline are modified. This project aims to conduct sensitivity analysis of supervised ML models. It will investigate which model properties are influenced by which changes in the underlying data and algorithmic pipeline. I will measure what factors in the models change and how, based on manipulating variables in the ML pipeline. I will evaluate the effects of either eliminating data points, or adding synthetic data points produced by generative models, and then observe the resulting impact on the target models. The research will be conducted using human-generated data from educational contexts. The results will include a systematic description of causal effects on bias under the defined conditions. This knowledge will improve the understanding of methodological choices for ML modeling.