Research

My research lies at the intersection of theoretical statistics, statistical learning, and machine learning, with a particular focus on applications involving imbalanced data. Specifically, I am interested in:

• establishing theoretical guarantees (e.g., central limit theorems with explicit convergence rates) for learning algorithms (random forests, bagged nearest neighbors, etc,..) trained on subsamples drawn without replacement;
• addressing class imbalance, a common challenge in industrial applications such as insurance, finance, and fraud detection;
• developing debiasing procedures based on the odds ratio, aimed at restoring the consistency of these algorithms under imbalance.

Recent works

I am interested in hybrid modelling of time series and neural networks. More specifically, I am interested in a hybrid modelling framework that integrates the linear structure of time series with the non-linear learning capabilities of deep neural networks, including RNNs, LSTMs and others.

Preprints

• Infinite random forests for imbalanced classification tasks , M. Mayala, O. Wintenberger, C. Tillier and C. Dombry
• Asymptotic Normality of Infinite Centered random Forests-Application to Imbalanced Classification. , M. Mayala, E. Scornet, C. Tillier and O. Wintenberger
• Mayala, M., (Ongoing). Imbalanced learning : Application to fraud detection.
• Mayala, M., and Gnandi, E. (Ongoing). When Random Forests Meet Time Series: hybridization.