Offre en lien avec l’Action/le Réseau : SIMDAC/– — –

Laboratoire/Entreprise : LIFO
Durée : 3 ans
Contact : Patrick.Marcel@univ-orleans.fr
Date limite de publication : 2025-02-26

Contexte :
De nombreux domaines nécessitent l’analyse de gros volumes de séquences de diverses complexités (en termes de périodicité, complétude, multivariée ou non, etc.) et en particulier de leur similarité. On peut citer les domaines aussi variés que le médical (e.g. stratification de patients, alignements de gènes), le social (analyse de trajectoires sémantiques), la science des données (génération et recommandation de pipelines d’exploration), etc.
Par exemple, le groupement de patients suivis sur de longues périodes peut être vu comme un problème de recherche et calcul de similarité sur des séquences complexes : les séquences sont apériodiques (la fréquence des rendez-vous médicaux n’étant pas fixe), multivariées (plusieurs informations sont enregistrées à chaque rendez-vous), incomplètes (les informations enregistrées peuvent varier d’un patient à l’autre).

Il est souvent nécessaire d’optimiser du calcul de similarités sur ces gros volumes de données de type séquences. Ce thème est à la croisée des domaines HPC (calcul haute performance) et analyse et exploration de données. Il recouvre différents challenges scientifiques : prise en compte de la spécificité des données (séquences, séries, trajectoires, etc.), définition de méthodes de réduction de dimensionalité et indexation, parallélisation des étapes du calcul de similarité, adaptation d’approches existantes (par exemple, sur séries temporelles ou données spatio-temporelles), etc.

Sujet :
Dans ce contexte, l’objectif de cette thèse est d’étudier différentes approches de recherche approximative pour le calcul de similarité de séquences complexes sur architecture HPC avec accélération GPU.
Parmi les approches, on s’intéressera plus particulièrement à utiliser et combiner des approches de Locality Sensitive Hashing (LSH), de réduction de dimensionalité, d’indexation, et d’échantillonage.

Profil du candidat :
Les candidats devront posséder un Master en informatique ou un niveau équivalent. Ils devront posséder un bon niveau en programmation, base de données, parallélisme et mathématique.

Formation et compétences requises :

Adresse d’emploi :
LIFO, Université d’Orléans

Document attaché : 202502261521_Sujet_de_th_se___approches_stochastiques_pour_le_calcul_de_similarit_s__de_s_quences_complexes (4).pdf

Offre en lien avec l’Action/le Réseau : – — –/Doctorants

Laboratoire/Entreprise : IMT Atlantique LabSTICC CNRS (Brest); AIRBUS DS (É
Durée : 3 years
Contact : cecile.bothorel@imt-atlantique.fr
Date limite de publication : 2025-03-20

Contexte :
Summary: This thesis focuses on anomaly detection, explanation, and labeling in complex networks. We would like to explore and to propose a new adaptive and hybrid explanation method that incorporates interactions with domain experts using preference models from the MCDA field.

Location: IMT Atlantique LabSTICC CNRS (Brest); AIRBUS DS (Élancourt)

Keywords: Complex Networks, Graphs Autoencoders, Anomaly Detection, Explainable AI (XAI), Multi-Criteria Decision Aiding (MCDA)

Supervisors : Cécile Bothorel, Lina Fahed, Arwa Khannoussi, Guillaume Gadek

Funding: COFUND SEED (Co-funded by the European Union) https://www.imt-atlantique.fr/en/research-innovation/phd/seed

Eligibility rules: (1) Did not spend more than 12 months in France since 20 March 2022 (last 36 months). (2) Is / will be awarded a master-level diploma or equivalent for Phd start (from September 2025) (he/she can be graduated during summer) and does not already have a doctoral degree.

To apply: https://seed-apply.imt-atlantique.fr

Application deadline: March 20, 2025

Detailed subject: https://www.imt-atlantique.fr/sites/default/files/recherche/doctorat/seed/research-topics/4-anomaly-detection.html

Starting date: fall 2025

For any question: please contact supervisors cecile.bothorel@imt-atlantique.fr & guillaume.gadek@airbus.com

Sujet :

1. Definition

Keywords: Complex Networks, Graphs Autoencoders, Anomaly Detection, Explainable AI (XAI), Multi-Criteria Decision Aiding (MCDA)
1.1. Domain and scientific/technical context

Temporal graphs, representing interactions over time, are crucial for analyzing datasets in areas like Industry 4.0, finance, transportation, biology, social networks, cybersecurity, and defence and intelligence. Detecting anomalies in temporal networks reveals unusual patterns and events, thus providing deep insights into the system behavior over time. Such graphs or networks typically grow every second and gather millions of attributed nodes and edges. Relevant behaviors are grounded in the nodes and edges characteristics as well as in higher-level patterns (local neighborhood, temporal similarities). Operational needs are based on constant monitoring (anomaly detection, alerts), for which there is a very strong need for tools: for the detection in itself, but also for the understanding of the detected anomalies in order to enable quick and relevant responses and preventive measures. Notions of traceability and actionability of the alert are also key to the adoption of the technology.
1.2. Scientific/technical challenges

This thesis focuses on anomaly detection, understanding, and labeling in complex networks for socially impactful applications such as social networks, financial exchange, health, defence, energy, etc. The two main challenges are: (i) the limited access to labeled data for anomaly detection, (ii) and when labels are obtained, they are often incorrect or unusable due to errors made by domain experts in labeling anomalies. To address these challenges, we propose to take advantage of three research areas: anomaly detection (for graphs), explainable AI (XAI), multi-criteria decision aiding (MCDA).

In order to detect anomalies, we will study the GNNs (Graph Neural Networks), and the use of auto-encoders de-signed for semi-supervised tasks with a small training set even if it contains labeling errors [1]. Several graph explanations methods have been proposed in the literature [2] that focus on different graph elements (nodes, edges, features). Both graph elements describing anomalies and explanations are criteria that experts can use to label anomalies. However, this may not provide actionable insights as experts may focus on intuitions derived from previous expertise. The challenge here is to provide experts with intuitive graph elements and explanations allowing to understand the anomalies.
1.3. Considered methods, targeted results and impacts

We propose a new adaptive and hybrid explanation method that incorporates interactions with experts. This can be done using preference models from the MCDA field, which allow the representation of decision strategies and human behaviour [3]. We expect to:

Provide explanations generated from traditional XAI methods [8] and a combination of dedicated eval-uation metrics.
Enrich and adapt explanations with multiple criteria related to multiple domain experts. Such criteria include the experts’ decision strategies, their behaviours, and insights into their prior expertise.
Iteratively involve experts in the loop, i.e., the interaction between the explanation method and the experts can be performed iteratively in such a way that at the end the experts are given the intuitive graph elements and explanations they need to understand the anomalies well and to label them correctly.

In this project, we plan to develop an experimental protocol on both synthetic and real-world impact datasets. This work will be an important step forward in the field of anomaly detection and understanding, and will open important perspectives related to the intersection of our different research domains.
1.4. Environment (partners, places, specific tools and hardware)

The academic partners are members of the DECIDE team at Lab-STICC (CNRS) and IMT Atlantique’s Data Science Department (DSD) in Brest, where interdisciplinary research exploit synergies between decision support and data science to address scientific, industrial and societal issues arising from decision-making problems in complex sys- tems (environment, transport, energy, social networks, health, defence).

The industrial partner, Airbus Defence and Space, is participating through its team of Artificial Intelligence for De- fence Digital. The team, based in Elancourt near Paris, is constituted of 20 data scientists, and contributes on re- search, technology development and deployment of AI assets within Airbus products, mainly in the Defence & In- telligence areas.

Airbus provides 3 use cases with datasets and interaction with business experts related to the use cases, all dir- ectly related to the Intelligence business. A) detecting coordinated behaviour in social networks for Cyber Inform- ation Warfare. B) highlighting patterns and edges of interest in communication interceptions (COMINT), most likely through simulated data. C) Smart assistant for investigation analyst on Knowledge Graphs: the product Massive Intelligence extracts & generates high-level data under the form of entities and relations, through the IKDB software. The tool would help the end-user to raise alerts on the extracted knowledge itself, highlighting suspicious cases and connections.
1.5. Interdisciplinarity aspects

The work combines 3 research domains: anomaly detection (for graphs), explainable AI (XAI), and multi-criteria decision aiding (MCDA). This thesis involves both theoretical, experimental and technical research to to serve the industrial interests and applications of Airbus Defence&Space.
1.6. References

[1] GILES, Bastien, JEUDY, Baptiste, LARGERON, Christine, et al. Suspicious: a Resilient Semi-Supervised Framework for Graph Fraud Detection. IEEE 35th International Conference on Tools with Artificial Intelligence (ICTAI), 2023.

[2] YUAN, Hao, YU, Haiyang, GUI, Shurui, et al. Explainability in graph neural networks: A taxonomic survey. IEEE transactions on pattern analysis and machine intelligence, 2022.

[3] KHANNOUSSI, Arwa, OLTEANU, Alexandru-Liviu, MEYER, Patrick, et al. A metaheuristic for inferring a ranking model based on multiple reference profiles. Annals of Mathematics and Artificial Intelligence, 2024.

[4] GADEK, Guillaume. “From community detection to topical, interactive group detection in Online Social Networks.” IEEE/WIC/ACM International Conference on Web Intelligence-Companion Volume. 2019.

[5] PRIEUR, Maxime, et al. “Shadowfax: Harnessing textual knowledge base population.” Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2024.

[6] BAUTISTA, Esteban, BRISSON, Laurent, BOTHOREL, Cécile, SMITS, Grégory. “MAD: Multi-Scale Anomaly Detection in Link Streams”. The 17th ACM International Conference on Web Search and Data Mining, Mar 2024, Mérida (Yucatan), Mexico.

[7] DAO, Vinh-Loc, BOTHOREL, Cécile, LENCA, Philippe. Community structure: A comparative evaluation of community detection methods. Network Science, 2020, 8 (1), pp.1-41.

[8] CHRAIBI-KAADOUD, Ikram , FAHED, Lina, LENCA, Philippe. Explainable AI: a narrative review at the crossroad of Knowledge Discovery, Knowledge Representation and Representation Learning. MRC@IJCAI 2021: Twelfth International Workshop Modelling and Reasoning in Context, 2021, pp.28-40.
2. Partners and study periods
2.1. Supervisors and study periods

IMT Atlantique: Prof.Cécile Bothorel, Assoc.-Prof. Lina Fahed and Assoc.-Prof. Arwa Khannoussi, IMT Atlantique, Brest, France.

Industrial partner: Dr. Guillaume Gadek, Airbus Defence and Space, Versailles, France

The PhD student will stay 9 months at Airbus Defence and Space.
Academic international partner(s): The PhD student will also spent 3 months at an international academic partner, probably LUT University, Finland (to be confirmed).

2.2. Hosting organizations
2.2.1. IMT Atlantique

IMT Atlantique, internationally recognized for the quality of its research, is a leading French technological university under the supervision of the Ministry of Industry and Digital Technology. IMT Atlantique maintains privileged relationships with major national and international industrial partners, as well as with a dense network of SMEs, start-ups, and innovation networks. With 290 permanent staff, 2,200 students, including 300 doctoral students, IMT Atlantique produces 1,000 publications each year and raises 18€ million in research funds.
2.2.2. Airbus Space and Defence

Airbus Space and Defence purpose is to improve life on Earth and beyond through our cutting-edge space technologies. From in-orbit delivery of satellites and spacecraft equipment to the smallest electronic components, Airbus provides products and services to customers around the world. We deliver telecommunications and navigation satellites that enable people to connect everywhere and navigate safely on Earth. The data from Airbus-built Earth observation satellites, such as Sentinel-2 or MetOp, bring insight that helps us to better understand and protect our planet.

Profil du candidat :

Formation et compétences requises :

Adresse d’emploi :
IMT Atlantique, campus Brest

Document attaché : 202502251404_4-anomaly-detection.pdf

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LIFAT & LIFO
Durée : 3 ans
Contact : alexandre.chanson@univ-tours.fr
Date limite de publication : 2025-05-12

Contexte :
In the era of data-driven decision-making, extracting insights from large datasets is crucial. Data narration refers to transforming data insights into interactive visual stories to enhance understanding and communication. While recent advances in AI and LLMs have introduced automation in data exploration and storytelling, challenges remain in personalization, user intent recognition, and interactive data narration.

*** Detailed subject attached ***

Sujet :
Key Research Questions :

– User Intent & Interaction: How can user preferences and feedback guide LLM-driven data storytelling?
– Personalization: How can data stories be adapted to different audience profiles, knowledge levels, and presentation styles?
– Exploration-Narration Interplay: How can data exploration and storytelling be seamlessly integrated to allow iterative user intervention?
– Quality Assessment: How can we evaluate and benchmark the effectiveness of generated data stories?

The candidate is expected to contribute to one of the first three research questions while considering the fourth as a transversal aspect.

*** Detailed subject attached ***

Profil du candidat :
Master’s degree in Computer Science :
– Strong background in databases and machine learning
– Interest in data exploration, storytelling, or NLP

Formation et compétences requises :
— Application Deadline: May 12, 2025 —
To apply, please email the following documents to the supervisors:
– CV
– Master’s transcripts
– Cover letter
– Reference letters, if any

Adresse d’emploi :
Université de Tours site de Blois

Document attaché : 202502191608_phd-llm-storytelling.pdf