Algorithmic fairness
Optimal-transport methods that correct discrimination in predictive models, for one sensitive attribute or several at once, with statistical guarantees on what accuracy that costs.
Fairness · Trustworthy AI · Insurance
Research
Statistical learning for settings where a prediction has consequences: fairness, calibration and labelling, mostly in insurance and finance.
Predictive models in insurance and finance are often trained on data that already encodes historical inequality, and then deployed in settings where the decision matters a great deal to the person on the other side. Two questions follow from that. First: given a model you already have, how do you correct its discriminatory behaviour, for one protected attribute or several at once, and what does the correction provably cost you in accuracy? Second: when labels are scarce and expensive, which observations should you pay to label, and how does that choice feed back into fairness?
Underneath both sits a third concern: whether the numbers a model produces mean anything. A score that is well ranked but badly calibrated is not a probability, and a fairness correction applied on top of it can quietly destroy what calibration there was. Probability calibration and Bayesian statistics are the tools I use for that part.
The methods are mostly optimal transport, Wasserstein barycenters, semi-supervised learning, calibration and Bayesian inference; the applications are motor and health insurance pricing, mortality scoring, ESG text analysis and epidemic early-warning. Work in the R&D AI Lab at Milliman extends this to generative AI, where the same questions reappear with different vocabulary.
Themes
Lines of work
Trustworthy AI for insurance
Pricing, scoring and mortality models that regulators and policyholders can actually scrutinise: interpretability, privacy and bias auditing on real actuarial portfolios.
Probability calibration
A score is only useful if its numbers mean what they claim. Recalibrating predicted probabilities so that they match observed frequencies, and keeping them calibrated once a fairness correction has been applied on top.
Bayesian statistics
Latent variable models, variational inference and MCMC, used both as a research tool for uncertainty quantification and as the backbone of the Bayesian machine learning course I teach.
NLP, LLMs and generative AI
Extracting structure from unstructured text: ESG concepts in corporate reporting, early signals of infectious disease outbreaks, and retrieval- augmented systems for insurance.
Semi-supervised & active learning
When labels are expensive, which ones should you buy? Sampling and labelling strategies that keep both cost and bias under control.
Publications
Papers & preprints
Filter
-
A sequentially fair mechanism for multiple sensitive attributes
AAAI 2024Covered by the Montreal AI Ethics Institute
A sequential Wasserstein-barycenter approach that mitigates unfairness across several sensitive attributes at once, with an interpretable fairness path.
-
Fairness guarantee in multi-class classification
Journal of Machine Learning Research (JMLR)
Optimal fair classifiers under demographic parity for multi-class problems, with statistical guarantees on the fairness/accuracy trade-off.
-
Mitigating discrimination in insurance with Wasserstein barycenters
BIAS 2023
How optimal-transport fairness corrections behave on real insurance pricing data, and what they cost in premium accuracy.
-
Fairness in multi-task learning via Wasserstein barycenters
ECML-PKDD 2023, Research Track
Extends optimal-transport post-processing for demographic parity to simultaneous regression and binary classification tasks.
-
Semi-supervised learning in insurance: fairness and labelling
PhD thesis, Institut Polytechnique de Paris (CREST-ENSAE)Awarded best actuarial thesis in France (SCOR Prize 2022)
A learning system for insurance that is accurate in prediction, cheap in labelling and ethical in transparency and fairness.
No publication matches that filter.
Talks
Selected presentations
-
March 2024
Algorithmic fairness for multiple sensitive attributes, with applications in insurance
Risk Forum
-
February 2024
A sequentially fair mechanism for multiple sensitive attributes
- 22 Sep 2023
- 21 Sep 2023
- 07 Feb 2023
-
15 Jun 2022
PhD defence: semi-supervised learning in insurance, fairness and labelling
Institut Polytechnique de Paris
- 28 Apr 2020
- 29 Nov 2019
Talk list last curated in March 2024. The CV has the complete record.
Collaborations
I supervise research internships and co-author with academic and industry teams. If your problem involves fairness, labelling or trustworthy models in a regulated setting, I would be glad to hear about it.