arXiv subjects

Jean-François Plante

Publications and source records attributed to Jean-François Plante.

4 recordsLinked to original sources

arXiv2026

Let's Unlearn Stereotypes Before Decision-Making: Assessing the Impact of Intrinsic Bias Mitigation on Downstream Fairness in LLMs

Large Language Models (LLMs) are increasingly used in high-stakes decision-making systems, where biased predictions can reinforce social and economic disparities. Although prior work has examined intrinsic representational bias and unfair downstream behavior separately, it remains unclear whether mitigating intrinsic bias leads to fairer downstream outcomes. We introduce Fairness-Aware Concept Unlearning (FACU), a model-level mitigation method that adapts concept unlearning to fairness-oriented representation balancing. Unlike suppression-based approaches, FACU explicitly regularizes probability differences between stereotypical and anti-stereotypical associations while preserving predictive performance and language modeling quality. We evaluate FACU across three open-source LLMs, multiple intrinsic bias benchmarks, and three socio-economic classification datasets using both frozen LLM embeddings and LoRA-fine-tuned classifiers. FACU produces statistically significant reductions in intrinsic gender bias that are associated with downstream fairness improvements across most evaluated settings, datasets, models, and fairness metrics, without significantly degrading predictive performance. Combining FACU with extrinsic mitigation methods, particularly counterfactual data augmentation, yields further fairness improvements. These findings suggest that fairness-aware intrinsic mitigation can support fairer LLM-based decision-making and that bias mitigation should be addressed across both model development and downstream deployment stages.

cs.CL↗

arXiv2025

A common zero-inflation bivariate Poisson model with comonotonic and counter-monotonic shocks

There are numerous applications which involve modeling multi-dimensional count data, notably in actuarial science and risk management. When such data exhibit an excess of zeros, common count models are no longer suitable. With multivariate data, characterizing an appropriate dependence structure is equally critical in order to adequately assess the underlying risk inherent in the joint counts. In this work we propose a new bivariate zero-inflated Poisson model appropriate for modeling pairs of counts with a surplus of zeros. The proposed construction is based on a mixture model involving a common mass at zero along with a Poisson random pair. Various forms of dependence are considered for the latent Poisson pair, allowing for both negative and positive dependence. Several model properties are explored, notably the joint probability mass function and implied dependence structure. The method of moments and maximum likelihood approaches are described and implemented for estimation. The usual asymptotic properties of the estimators are derived, and their finite sample properties are explored through extensive simulations. The practical use of the proposed model is further illustrated through two real data applications.

stat.ME↗

arXiv2021

Fully Bayesian inference for spatiotemporal data with the multi-resolution approximation

Large spatiotemporal datasets are a challenge for conventional Bayesian models because of the cubic computational complexity of the algorithms for obtaining the Cholesky decomposition of the covariance matrix in the multivariate normal density. Moreover, standard numerical algorithms for posterior estimation, such as Markov Chain Monte Carlo (MCMC), are intractable in this context, as they require thousands, if not millions, of costly likelihood evaluations. To overcome those limitations, we propose IS-MRA (Importance sampling - Multi-Resolution Approximation), which takes advantage of the sparse inverse covariance structure produced by the Multi-Resolution Approximation (MRA) approach. IS-MRA is fully Bayesian and facilitates the approximation of the hyperparameter marginal posterior distributions. We apply IS-MRA to large MODIS Level 3 Land Surface Temperature (LST) datasets, sampled between May 18 and May 31, 2012 in the western part of the state of Maharashtra, India. We find that IS-MRA can produce realistic prediction surfaces over regions where concentrated missingness, caused by sizable cloud cover, is observed. Through a validation analysis and simulation study, we also find that predictions tend to be very accurate.

stat.CO↗

arXiv2016

Statistical Inference, Learning and Models in Big Data

The need for new methods to deal with big data is a common theme in most scientific fields, although its definition tends to vary with the context. Statistical ideas are an essential part of this, and as a partial response, a thematic program on statistical inference, learning, and models in big data was held in 2015 in Canada, under the general direction of the Canadian Statistical Sciences Institute, with major funding from, and most activities located at, the Fields Institute for Research in Mathematical Sciences. This paper gives an overview of the topics covered, describing challenges and strategies that seem common to many different areas of application, and including some examples of applications to make these challenges and strategies more concrete.

stat.ML↗