PhD candidate, Machine Learning
SerVal group, SnT, University of Luxembourg
Luxembourg
I'm a PhD candidate in Machine Learning at the SnT of the University of Luxembourg, supervised by Prof. Yves Le Traon and Dr. Sylvain Kubler. My research focuses on synthetic data generation for time series: producing, selecting, and evaluating augmented samples that help downstream models when real data is scarce, imbalanced, or drawn from a different distribution.
Many futures are possible. I'd be glad if one of them has us collaborating.
Mistral AI's new forecasting models, all four sizes, on series a person could forecast by hand, scored against the best forecast possible: lines, exponentials, cycles, breaks, covariates and long horizons, next to Chronos-2 and TimesFM 2.5.
ASCENSION shapes a VAE latent space with a contrastive loss, then generates new samples with an α-scaling mechanism that progressively expands each class's posterior covariance while preserving class identity. Evaluated on the 102 univariate UCR datasets with two deep classifiers and a time-series foundation model against eight augmentation methods, with the largest gains where training data is scarcest.
A multi-agent system that automates the classification and tagging of predictive-maintenance datasets, lowering the curation effort for downstream learning pipelines.
A Wasserstein CycleGAN with gradient penalty (W-CycleGAN-GP) generates synthetic samples in an unlabeled target domain from labeled source-domain samples. Combined with adversarial and correlation-alignment domain-adaptation losses, the augmented data improves remaining-useful-life prediction on NASA's C-MAPSS turbofan benchmark.
BALANCER is a machine-learning framework that recommends the most effective augmentation technique for an imbalanced time-series classification task. Validated on an empirical study of 720 datasets and explained with SHAP-based interpretability.
If you're a student interested in time series, augmentation, or synthetic data generation, feel free to reach out.