Artificial intelligence for data science and cybersecurity

Researchers in this theme study new models and algorithms in the field of data integration, classification and (co)-clustering. They also develop theoretical and applied research in the field of cyber security and resource management in cyber-physical systems

Scientific referent

Coordinator: Mohamed Nadif

Scientific leaders:

  • Severine Affeldt
  • Lazhar Labiod
  • Ahmed Mehaoua
  • Osman Salem

Presentation of the thematic area

  •  Machine learning for data science is an essential field of study in artificial intelligence. It comes in different forms: unsupervised, semi-supervised, supervised by reinforcement. Although multiple algorithms, models and strategies are available today, many major challenges still remain, in many different domains.
  •  The researchers of the team focus on several issues related to learning including (Co)-clustering and dimensionality reduction (unsupervised/semi-supervised learning) as well as supervised classification. Our work is based on different approaches such as matrix factorization, mixture models, latent block models, spectral decomposition and deep learning. Our main objective is to propose innovative models and algorithms that are efficient and easily exploitable in practice. Thus, the methods we propose are dedicated to the processing of multi-source data of different natures with applications in various domains such as textual data analysis, automatic natural language processing, bioinformatics, collaborative filtering, mediation analysis and cyber security. In addition, we are interested in the medical field, developing new machine learning methods and user-friendly software that can integrate various types of omics data to identify the players in complex human diseases.
  • The researchers of the group are also developing theoretical and applied research in the area of cyber security and resource management in cyber-physical systems in particular anomaly detection for wireless medical body sensor networks. The group's contributions are oriented towards the design, optimization and performance evaluation of new protocols, algorithms, tools and formal models. They allow, thus, to provide quality, and security of communications and data, in next generation health physics systems such as chronic disease detection (Ischemia, epilepsy, etc.).

Keywords

Machine and Deep Learning ; Co-clustering ; Factorization ; Spectral Clustering ; Mixture models ; Attributed Network Embedding ; Mediation analysis ; Wireless Sensor Networks ; Internet of Medical Things ;Security and Anomaly detection ; Resource Optimization

 

Topics covered

  •  Spectral clustering via ensemble deep autoencoder learning and evaluation on image data.
  • Regularized bi-directional co-clustering for biomedical texts.
  • Endotypes identified by cluster analysis in asthmatics and non-asthmatics and their clinical characteristics at follow-up: the case-control EGEA study.
  • Unsupervised text mining for assessing and augmenting GWAS results. 
  • Real-time biomedical data analysis systems based on Machine Learning (ML) and Wireless Body Sensor Networks (WBAN),
  • Sensor-based remote health monitoring, Sensor-based Human activity recognition, Sensor-based Ischemia and Epilepsy detection, 
  • Cybersecurity threats Detection using AI/ML, Blockchain-based Anomaly and threats detection for Internet of Things (IOT).

Applications

(Co)-clustering de documents textuels

(Co)-clustering d'images

Softwares and Packages

L'algorithme BCOT concerne le biclustering basé sur le transport optimal.

Article de référence : C. Fettal, L. Labiod, M. Nadif. Efficient and effective optimal transport-based biclustering. NeurIPS, 32989-33000, 2022

L'algorithme SC3 est dédié au “subspace convolutional co-clustering".

Article de référence : Boosting subspace co-clustering via bilateral graph convolution. IEEE Transactions on Knowledge and Data Engineering, 36(3): 960-971, 2024

L'algorithme SAGSC est dédié au “subspace clustering”.

Article de référence : C. Fettal, L. Labiod, M. Nadif. Scalable attributed-graph subspace clustering. AAAI, 7559-7567, 2023.

L'algorithme LMGEC est dédié au Multi-vues des graphes attribués.

Article de référence : C. Fettal, L. Labiod, M. Nadif. Simultaneous linear multi-view attributed graph representation learning and clustering. WSDM, 303-311, 2023.

Le package Caeclust implémente une méthode s’appuyant sur un algorithme de clustering de type deep spectral. 

Ce travail a donné lieu à une publication. S. Affeldt, L. Labiod, M. Nadif. CAEclust: A consensus of autoencoders representations for clustering. IPOL. 590-603, 2022. Ce travail est issu d’un travail des mêmes auteurs dans Pattern Recognition journal (2020).

L'algorithme GCC est dédié au “graph convolutional clustering”.

Article de référence : C. Fettal, L. Labiod, M. Nadif. Efficient Graph Convolution for Joint Node Representation Learning and Clustering. WSDM, 289-297, 2022.

Le package python TensorClus est dédié au clustering et co-clustering de données tensorielles.

Article de référence : R. Boutalbi, L. Labiod, M. Nadif. TensorClus: A Python library for tensor (co)-clustering. Neurocomputing, 464- 468, 2022.

WordGraph est un package python permettant de reconstruire des modèles graphiques causaux interactifs à partir de données textuelles. Il a été publié à WSDM 2024 et a reçu le prix du meilleur papier “software” présenté à EGC 2024

CORPEX est une interface d'aide à l'analyse de corpus via la visualisation interactive de co-clusters pour soutenir l'exploration de thèmes pour un ensemble de textes. 

Cette interface a été publiée dans EGC 2023 par A. Ferdjaoui, Amira Tlati, S. Affeldt, M. Nadif; CORPEX : Analyse exploratoire d'un corpus biomédical à l'aide de la classification croisée.

Le package  dcblockmodels  implémente des algorithmes de co-clustering pour les données de comptage basés sur le modèle de blocs latents (LBM). Il propose deux modèles principaux : un LBM dynamique (dLBM) pour les données représentées sous forme de séries de matrices d'adjacence, et un LBM semi-supervisé (ou contraint) (HLBM) utilisant des contraintes par paires dans les espaces des lignes et des colonnes.

Pour plus de détails : 

Le package ELBMcoclust implémente plusieurs algorithmes de co-clustering. 

Article de référence :  S. Hoseinipour, M. Aminghafari, A. Mohammadpour and M. Nadif “A Sparse  Exponential Family Latent Block Model for Co-clustering”. Advances in Data Analysis and Classification, 1-37.

Main publications

Projects

  • Project ANR GePhEx (S. Affeld, 2019) : Learning causal effects between phenome and exposome from large amounts of heterogeneous data in human complex diseases.
  • Project ANSES MOLDASTH (R. Nadif, 2021) Moulds in dwellings, inflammation, immune response, and ASTHma endotypes in the CONSTANCES cohort.
  • Project Emergence Idex Spectrans (M. Nadif, 2021). Specialised corpora and neural translation.
  • Project CDC Informatique. (M. Nadif) Detecting anomalies and reversals in finance.
  • Project  THALES (M. Nadif, 2019) Hybridization of AI algorithms with business knowledge for rail transportation.

  • Project SOPRA-STERIA and AIRBUS-APSYS. Security of Industrial Internet of Things based on Blockchain.

  • Project ORANGE LABS. Real-time Network Service Detection, Classification and Analysis from encrypted real-time traffic communications

Interactions with the other themes of  Centre Borelli