My research focuses on reliable and trustworthy statistical learning at the intersection of machine learning, statistics, and causality. I develop methods that combine predictive performance, interpretability, and robustness under real-world constraints such as heterogeneous, imperfect, and weakly labeled data. My work ranges from interpretable, knowledge-driven models to deep learning approaches, and more recently explores gray-box AI to bridge black-box learning with causal transparency. A central theme of my research is the integration of deep learning and causality; including causal representation learning, counterfactual robustness, causal inference from longitudinal data, and causal discovery; to design intrinsically interpretable models for reliable decision-making.

In my postdoc, I proposed an ensemble method for uncertain measures. This method is based on smooth regression and decision trees which are constructed on a particular learning set. Each quantitative observation of the learning set is assumed to be an uncertain measure, i.e., has a known distribution (then a known variance parameter). To deal with this precious information, the empirical risk to minimize takes into account the probability that each observation belongs to each region of the input space. I worked on encoding the underlying algorithm for one smooth decision tree and random forest construction. The resulting approach called Smooth And Consistent Probabilistic Regression Trees was published in NeurIPS 2020.

The perspectives are to extend it to the classification task. Indeed, TotalEnergie company (financer of the project) is very interested in this work to learn regression or classification models for predicting welding quality. Therefore, I had the pleasure of working with welding experts on a real data set. A last perspective is to A recent perspective is to go beyond the association level to tackle causal questions from data which is often a key to discover how to alter and optimize the system of welding quality prediction.

In my Ph.D. thesis, I developped an EM simultaneous estimation of a structural equation modeling and its latent factors. To construct practically a model and identify the explanatory blocks of variables, I have proposed to use a PCA. Instead of this manual step, we propose to add an algorithm step which, given an assumed number of explanatory blocks, computes the probability that each explanatory variable belongs to one block. We are comparing the performances of two approaches: SEM-Gibbs (Stochastic Expectation Maximization) and Lasso.

Among several perspective of my Ph.D work, I am currently working on Causality and the great contribution that could belong.

• Modeling in machine learning
• Causal inference from longitudinal data
• Causal representation learning
• Robustness and trustworthy AI
• AI Interpretability
• AI Explainability
• Multi-modal learning
• Ensemble learning
• Heterogeneous, complex and multi-modal data
• Weakly supervised and few-shot learning
• Domain Adaptation
• Structural equation modeling with latent factors or latent variables
• EM algorithm
• Oncology data
• Health related quality of life