Machine learning models are increasingly trained on data integrated from multiple sources or levels. For example, in healthcare, a predictive model may combine patient-level clinical records , genomic data , and population-level socioeconomic indicators .
Data fusion can occur at different levels of the machine learning pipeline—namely, early fusion (feature-level), middle fusion (embedding or representation-level), and late fusion (decision-level). Each of these integration strategies may affect fairness differently. Data fusion is extensively studied in the context of improving predictive performance, yet its implications for algorithmic fairness remain underexplored.
While such multi-level data fusion can improve predictive performance, it raises important questions about fairness : when data are combined across levels, how are biases propagated or amplified, and how does the fairness of the resulting model compare to models trained on each source individually?
This project aims to:
- Survey existing fairness notions in the context of multi-level or multi-source data, including group fairness, intersectional fairness, and individual fairness.
- Develop experimental setups to evaluate how combining data from different levels affects fairness metrics across various protected attributes (eg, gender, ethnicity, socioeconomic status).
- Analyze trade-offs between predictive performance and fairness in fused models, and explore strategies to mitigate fairness degradation during data integration.
- Compare how fairness outcomes vary across fusion strategies using different types of machine learning methods, and provide a flexible framework for evaluating fusion approaches across models (eg, kernel methods or neural networks).
The work will provide both theoretical insights into the impact of data fusion on fairness and practical guidance for building equitable machine learning systems in domains where multi-level data are common, such as healthcare, social science, and genomics.
If you want to learn more or get involved, please contact Pooya Zakeri or Anne-Marie George.