Current Open Topics

This page presents a selection of current research directions and open methodological questions that I am actively developing. ## EEG clustering through tensor decomposition

EEG data are naturally multi-dimensional, combining subjects, channels, time, frequencies and experimental conditions. Tensor decomposition offers a way to preserve this multiway structure, but the connection between decomposition quality and clustering performance remains an open methodological issue.

Research question: How can tensor decomposition methods be designed or regularized to obtain meaningful and stable clusters of EEG profiles?

Clustering individual networks

In several biomedical and cognitive applications, each individual can be represented by a subject-specific network. The methodological challenge is different from classical community detection: the goal is not to identify communities within a single network, but to cluster individuals according to the structure of their networks.

Research question: How can subject-specific networks be compared and clustered while preserving interpretability and avoiding circularity in the estimation process?

Bayesian ensemble methods for clustering and prediction

Ensemble methods are powerful tools for improving stability and predictive performance, but their uncertainty is often difficult to quantify and communicate. Bayesian and bootstrap-based approaches provide a promising framework to combine robustness, uncertainty assessment and interpretability.

Research question: How can Bayesian ensemble methods produce stable predictions or partitions while providing meaningful uncertainty information?

Robust rank-based clustering

Many clustering problems involve data structures for which classical metric assumptions are too restrictive or unstable. Rank-based information offers a flexible way to compare observations or complex objects while reducing the impact of scale, noise and distributional assumptions.

Research question: How can rank-based similarity measures be used to build robust clustering procedures for complex and heterogeneous data?