Federated & edge learning
Training and optimisation of models in resource-constrained environments (Edge, IoT, embedded); federated learning under resource heterogeneity; knowledge distillation.

Ph.D. · Postdoctoral Researcher · CEA
I’m Mehdi Acheli, an AI researcher following intelligence from event logs to the edge and IoT, where data is rich and resources are scarce.
$ I am an AI researcher with a PhD in computer science, driven by research and innovation in data science.
My doctoral work at Université Paris-Dauphine-PSL introduced me to process science. I developed methods to discover behavioral patterns in the event logs of flexible processes, where traditional process models fail to capture how work actually unfolds. I then moved to industry R&D, first predicting process behavior at Logpickr, then at Sogeti, part of Capgemini, where I led a research project on fraud detection: identifying anomalies in financial processes from historical event-based data.
Returning to academia as a postdoctoral researcher at Télécom SudParis, where I supervised PhD students and interns, I turned to the Cloud-Edge-IoT continuum. There, too, processes are central: a service composition is itself a process, coordinating services with dependencies and execution order. My work followed this process through its lifecycle across cloud, edge and IoT devices: generating compositions from user intent with large language models, monitoring them for anomalies as they run, and using language models again to renegotiate service-level agreements when anomalies occur. Monitoring the continuum also means running learning models close to the devices themselves, which led me to work on training models on resource-constrained devices at its lower layers. Beyond the continuum, I worked with a cybersecurity firm on predicting security ratings from companies’ rating histories, another form of event-based historical data, and drew on my experience with language models to automate third-party risk assessment questionnaires.
Today, as a postdoctoral researcher at CEA, I focus on adapting learning models to the edge and IoT. This line of work, which began at Télécom SudParis with lightweight anomaly detection through knowledge distillation and a dual-model training approach, now continues with my research on federated learning across devices with heterogeneous resources.
Training and optimisation of models in resource-constrained environments (Edge, IoT, embedded); federated learning under resource heterogeneity; knowledge distillation.
Anomaly detection, orchestration process, its monitoring and adjustment, and agentic AI for the Cloud-Edge-IoT continuum.
Mining event logs, behavioural pattern extraction, causal inference, and next-activity prediction inspired by NLP and computer vision.
Generative graph diffusion for anomaly detection, graph neural networks for fraud detection and predictive process monitoring, and reinforcement-learning-based neural architecture search (RL-GraphNAS).
Security rating systems and supplier risk assessment.
CEA
Federated learning under resource heterogeneity.
Télécom SudParis
Sogeti, part of Capgemini
Logpickr, iGraphx
Predicting consequent activities in a process, drawing on NLP and computer vision.
No publication matches these filters.
DiCEI: Distilled iTransformer for Effective Lightweight Anomaly Detection Across the Cloud-Edge-IoT Continuum
IEEE International Conference on Cloud Engineering (IC2E), 2026. Accepted.
Dual Model Learning for Efficient and Effective Visual Services
IEEE International Conference on Web Services (ICWS), Sydney, Australia, 2026, pp. 771–781. DOI ↗
Semantic Labeling for Third-Party Cybersecurity Risk Assessment: A Semi-Supervised Approach to Intent-Aware Question Retrieval
International Conference on Knowledge-Based and Intelligent Information & Engineering Systems (KES), 2026. Accepted. arXiv ↗
Reliable Remediation Impact Prediction for Black-Box Security Ratings
arXiv:2607.16357, 2026. arXiv ↗
From Natural Language to TOSCA: Leveraging LLMs for Automated Service Composition
International Conference on Service-Oriented Computing, Springer, pp. 3–18, Dec. 2025. DOI ↗
Predictive Process Monitoring Using Object-Centric Graph Embeddings
Service-Oriented Computing – ICSOC 2024 Workshops, LNCS vol. 15833, Springer. DOI ↗
Mining a Minimal Set of Behavioral Patterns using Incremental Evaluation
arXiv:2402.02921, 2024. arXiv ↗
Discovering and Analyzing Contextual Behavioral Patterns from Event Logs
IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 12, pp. 5708–5721. DOI ↗
Behavioral Pattern Mining for Flexible Processes
Doctoral dissertation, Université Paris sciences et lettres. Supervised by Pr. Daniela Grigori. HAL ↗
Applying the Method of Reflections through an Event Log for Evidence-based Process Innovation
International Conference on Process Mining (ICPM), Aachen, Germany, pp. 105–112. DOI ↗
Efficient Discovery of Compact Maximal Behavioral Patterns from Event Logs
Advanced Information Systems Engineering (CAiSE 2019), LNCS vol. 11483, Springer. DOI ↗
About 230 hours of lab and tutorial teaching (TD/TP) at Université Paris-Dauphine-PSL during and after my Ph.D.
| Year | Course Unit | Type | Volume |
|---|---|---|---|
| 2021–2022 | Bases de Données Relationnelles | TD/TP | 52,5 h |
| 2020–2021 | Systèmes d’Exploitation | TD/TP | 40,5 h |
| C++ | TD/TP | 21,94 h | |
| 2019–2020 | Outils en Informatique | TP | 24 h |
| Bases de Données Relationnelles | TD/TP | 16,13 h | |
| Programmation C | TD/TP | 10,5 h | |
| 2017–2018 | Outils en Informatique | TP | 12 h |
| Bases de Données Relationnelles | TD/TP | 52,5 h |
Data science for resource monitoring, SLA renegotiation and anomaly detection in the Cloud-Edge-IoT continuum
Anomaly detection and monitoring in the Cloud-Edge-IoT continuum: from data models to intelligent detection and optimal workload placement
Contextual rating system for the automated assessment of IT infrastructures, integrating AI models
Automated orchestration across the Cloud-Edge-IoT continuum
Master II Vision et Machine Intelligente, Université Paris Cité
Exploring neural networks with modular architecture in the context of bank fraud detection
Master 2 Machine Learning for Data Science, Université Paris Cité
Financial fraud detection with Graph Neural Network methods
Master Professionnel en Informatique, Horizon School of Digital Technologies
Anomaly detection in dynamic graphs using generative graph diffusion
Université Paris-Dauphine–PSL
Data mining on business-process logs; behavioural pattern extraction and feature engineering; data analytics, correlation studies and causal inference. Taught C, C++, Unix, RDBMS, HTML, MS Access.
Thesis: Behavioral Pattern Mining for Flexible Processes, supervised by Pr. Daniela Grigori. Read thesis ↗
Higher National School of Computer Science | LIAS (ENSMA)
Graduation project: benchmarking of databases (NoSQL, distributed RDBMS, Hive) and development of data and query generators. Public outreach through tutorials.
Interested in collaborating, supervising, or reviewing? Let’s talk.