PhD student at LISN, CentraleSupélec / Université Paris-Saclay
I work on self-supervised learning and anomaly detection for tabular data, with a focus on uncertainty quantification. Code for the recent papers is linked below.
Self-supervised learning for tables. Tabular data has no natural augmentations, so methods built on crops and flips do not transfer. T-JEPA predicts the latent representation of one feature subset from another, without reconstructing the input. ICLR 2025.
Convergence collapse in anomaly detection. Autoencoders, Deep SVDD, RDP and flow-matching detectors all score anomalies as a residual against a fixed target. As training converges that residual shrinks on anomalies too, so detection degrades while the training loss keeps falling. Early stopping does not fix it, since the best epoch moves by two orders of magnitude across seeds. KCM replaces the trained map with a closed-form kernel smoother and has no training phase at all. CIKM 2026, oral.
Uncertainty on tabular foundation models. TFMs are accurate on tabular data but poorly calibrated, measured across the 112 TALENT datasets. Masking a TFM's attention diagonal recovers all n leave-one-out conformal scores in a single forward pass. MICCAI 2026 MultiTab, oral.
Bold = me. Full list and talks at josecosta.fr.
| Paper | Venue | |
|---|---|---|
| Mitigating Convergence Collapse in Fixed-Target Anomaly Detectors via Kernel-Anchored Locality Regularization De Melo Costa, Popineau, Rimmel, Doan |
CIKM 2026, oral | code pypi |
| Single-Pass Conformal Cross-Modal Anomaly Screening with Tabular Foundation Models De Melo Costa, Ahn, Popineau, Rimmel, Doan |
MICCAI 2026 MultiTab, oral | code |
| Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets Ahn, Leite, De Melo Costa, Popineau, Doan, Rimmel |
CIKM 2026 | arxiv |
| Drive vs. Decay: Early-Training Stability and Identity-Biased Predictors for Joint-Embedding Predictive Architectures De Melo Costa, Ahn, Popineau, Rimmel, Doan |
Under review, 2026 | |
| High Performance, Low Reliability: Uncertainty Benchmarking for Tabular Foundation Models De Melo Costa, Popineau, Rimmel, Doan |
ESANN 2026 | doi arxiv code |
| Leveraging Self-Supervised Learning for Fraud Detection in Tabular Data De Melo Costa, Popineau, Rimmel, Doan |
19th Financial Risks Int. Forum | pdf |
| T-JEPA: Augmentation-Free Self-Supervised Learning for Tabular Data Thimonier*, De Melo Costa*, Popineau, Rimmel, Doan · *joint first authors |
ICLR 2025 | openreview arxiv code |
t-jepa  |
Official ICLR 2025 implementation, including the regularization tokens that keep JEPA training stable on structured inputs. |
kernel-contraction-matching  |
The CIKM 2026 artifact, packaged as a scikit-learn estimator. |
| high-performance-low-reliability | Uncertainty benchmark for tabular foundation models on the 112 TALENT datasets. |
| single-pass-conformal-calibration | Leave-one-out conformal calibration in one forward pass, for cross-modal anomaly screening. |
pip install kar-kcmfrom karkcm import make_kcm_pipeline
detector = make_kcm_pipeline().fit(X_train)
detector.anomaly_score(X_test) # higher means more anomalous- CIKM 2026, Rome, November. Presenting the anomaly detection work.
- MICCAI 2026, Strasbourg, September. Presenting the conformal screening work.
- Starting on sensitivity analysis for causal adjustment.
- Teaching assistant in algorithms and programming, and supervising third-year student projects on fraud detection for the LUSIS chair.
Python, PyTorch, scikit-learn, LaTeX, Slurm on Jean Zay. Portuguese (native), French (C1), English (C1).
Reach me at costa@lisn.fr.
Updated August 2026