Files
Traore_46631900_2024.pdf
Embargoed access - Adobe PDF
- 2.24 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- The topic of this Master's thesis is fairness in supervised machine learning. The work centres around the Datagotchi project, an electoral compass which aims to predict individuals' voting intentions based on their lifestyle and socio-demographic factors. The underlying method is a classification algorithm trained on a labelled dataset. This application is prone to issues of fairness for diverse reasons. Most notably, users are asked to input sensitive information such as their gender, sexual orientation and age, which must be handled with care. Recognizing bias toward sensitive demographics is essential to prevent misleading groups by giving inaccurate political recommendations. Additionally, ensuring fairness in Datagotchi is vital for creating an application that appeals to a broad audience. This work reviews definitions of fairness in machine learning and reports the most sensitive ones for our study case. Several methods have been designed and implemented to address concerns about fairness. The thesis approaches the issue of fairness from two perspectives. As a starting point, we discuss the subject from the traditional framework, aiming to enhance equity based on individuals' sensitive features. Subsequently, this work departs from the standard fairness notions. Indeed, the focus is also on developing a classification model that achieves fairness concerning the target classes. This requires defining a class-oriented fairness metric. For this part, the scope of the proposed definitions and methods extends beyond the study case presented here.
Citations
Traore, L. (2024). Group and class-oriented fairness in classification: a case study of Datagotchi. https://dial-mem.test.bib.ucl.ac.be/handle/2078.2/39689