Research & Publications

The Violation Situation Pattern: A Knowledge-Graph Pattern for Compliance Violations Under Review

EKAW 2026  ·  Preprint: arXiv:2606.03326  ·  Also on HAL

Proposes the Violation Situation Pattern (VSP), a knowledge-graph ontology pattern that reifies each compliance violation as a persistent graph object with a rule identifier, lifecycle state, and PROV-O provenance trail — so audit history becomes a graph traversal rather than a ticket search. Instantiated in a legal entity and contract lifecycle property graph, operationalizing four deontic rules (unauthorized signatures, expired mandates, missing confidentiality clause, missing breach-notification clause) through an FCL→Cypher→MERGE pipeline.

Central finding is rule-body independence: extending a rule's detection logic raises F1 from 0.312 to 0.602 on GDPRhub enforcement decisions while the pattern's identity, lifecycle, and evidence semantics stay the same — separating the pattern contribution from the detector contribution so detection logic can evolve without invalidating accumulated audit history.

Beyond a Single Test Year: A Temporally Honest AutoML Benchmark for MODIS Active-Fire Type Classification In Preparation

Manuscript in preparation  ·  Sept. 2025 – Present

A temporally honest AutoML benchmark that reconstructs the MODIS MCD14ML active-fire type attribute — absent from near-real-time feeds — from per-detection observables alone, an attribute prior literature has almost never used as a supervised learning target. Compares six learners (RF, XGBoost, CatBoost, LightGBM, MLP, Kolmogorov–Arnold Network) across eight class-imbalance strategies on 300K+ detections using Leave-One-Year-Out cross-validation, with Optuna hyperparameter search and statistical separation via Cochran's Q + Bonferroni-McNemar tests.

Three complementary case studies: a Mediterranean basin polygon (228K detections, four-class), a country-scale Türkiye corpus (71K detections, binary), and the same polygon repartitioned by COVID-19 PHEIC boundaries. Reports a candid negative result for KAN on structured satellite tabular data, with fully reproducible SHAP analyses and bootstrap CIs.

Neurosymbolic Reasoning with RNNLogic Manuscript

CentraleSupelec Big Data Research Project  ·  Oct. 2025 – Feb. 2026
Supervised by Laura Forero Camacho  ·  Co-authored with Olha Baliasina & Adrian Patricio

Enhances the RNNLogic neuro-symbolic framework's EM training with a reinforcement-learning-inspired rule sampler: a learned predictor scores candidate logical rules each iteration so the generator progressively trains on higher-quality rules rather than uniformly sampled ones. Extends the rule generator (GRU) with knowledge-graph clustering for more robust rule mining, and benchmarks on standard KG completion tasks.

Exploring Temporal Machine Learning Approaches for Predicting Methane Discharge in Atmospheric Studies Published

ACDSA 2025  ·  Aug. 7–9, 2025, Antalya  ·  IEEE Xplore

Benchmarks traditional ML and deep learning architectures — Random Forest, XGBoost, MLP, and Temporal Convolutional Networks — for spatiotemporal methane discharge prediction using multivariate continuous atmospheric time-series from the Alberta Air Data Warehouse. Validates predictive accuracy and computational efficiency across models, contributing a reproducible pipeline for affordable, scalable methane emission monitoring to support climate policy decisions.

Capitalizing the Predictive Potential of ML to Detect Fire Types Using NASA's MODIS Satellite Data for the Mediterranean Basin Published

ICAAI 2023  ·  Oct. 13–15, 2023, Istanbul  ·  ACM Digital Library

Investigates ML approaches for classifying wildfire types across the Mediterranean Basin using NASA FIRMS MODIS satellite data spanning 2019–2022. XGBoost achieves an overall F1 score surpassing 95% and an 84% macro F1 across varied fire event types — outperforming Random Forest — with validation on held-out 2022 data. The Mediterranean-Basin polygon framing and multi-class fire-type classification represent novel contributions; the MODIS type attribute had not previously been used as a supervised learning target in the literature.