Seydina Ousmane NIANG

INRIA Rennes — LACODAM

Rennes, France

I completed my PhD in the Maasai project-team (Inria), at the Laboratoire J.A. Dieudonné, Université Nice Côte d’Azur, under the supervision of Prof. Charles Bouveyron, Dr. Marco Corneli, and Prof. Pierre Latouche. I defended my thesis on 2 October 2026 — [slides].

My PhD research focused on deep generative models for the joint analysis of networks and continuous data. More broadly, my research interests lie at the intersection of probabilistic modelling, statistical learning, and machine learning, with a particular emphasis on graph-structured data, deep generative models (variational autoencoders, latent variable models), probabilistic inference, and applied statistics.

I am now a postdoctoral researcher at INRIA Rennes in the LACODAM team.

Feel free to reach out by email for any question related to my research.

news

Nov 30, 2025	Invited talk at the 3IA Côte d’Azur Postdoctoral Seminar, Nice, France — The Deep Latent Position Block Model for the Clustering of Nodes in Multi-Graphs. [Slides available upon request.]
Oct 31, 2025	Poster presentation at Sophia Summit 2025, Sophia Antipolis, France — The Deep Zero-Inflated Latent Position Block Model for the Clustering of Nodes in Graphs. [Poster available upon request.]
Jul 01, 2025	Poster presentation at the Machine Learning Summer School, Senegal — Importance Weighted Directed Graph Variational Auto-Encoder for Block Modelling of Complex Networks. [Poster available upon request.]
May 31, 2025	Oral presentation at the 56th Journées de Statistiques de la SFdS (JDS 2025), Marseille, France — The Deep Zero-Inflated Latent Position Block Model for the Clustering of Nodes in Graphs. [Slides available upon request.]
Oct 31, 2024	Poster presentation at Sophia Summit 2024, Sophia Antipolis, France — Conditional Denoising Diffusion Probabilistic Models for the Clustering of Images. [Poster available upon request.]

selected publications

The Deep Latent Position Block Model for the Clustering of Nodes in Multi-graphs

Seydina Ousmane Niang, Charles Bouveyron, Marco Corneli, and 2 more authors

Journal of Classification, Mar 2026

Published online: 12 March 2026

Abs DOI HAL Bib HTML PDF

Network data capture relationships among actors across multiple contexts, often forming clusters of individuals. These relationships frequently involve multiple types of interactions, requiring multidimensional networks, or multi-graphs, to fully capture their complexity. Latent position models (LPMs) embed nodes based on connection probabilities but cannot uncover heterogeneous clustering structures such as disassortative patterns. Stochastic block models (SBMs) excel at clustering but lack interpretable latent representations. To address these limitations, we introduce Deep-LPBMM, extending the deep latent position block model to multidimensional networks using a deep variational autoencoder with graph convolutional networks across multiple interaction types, enabling both effective clustering and enhanced visualization.
@article{niang2026deeplpbmm, title = {The Deep Latent Position Block Model for the Clustering of Nodes in Multi-graphs}, author = {Niang, Seydina Ousmane and Bouveyron, Charles and Corneli, Marco and Latouche, Pierre and Boutin, R{\'e}mi}, journal = {Journal of Classification}, year = {2026}, month = mar, doi = {10.1007/s00357-026-09541-w}, keywords = {Stochastic block model, Latent position model, Clustering, Multidimensional network analysis, Multi-graphs, Variational autoencoder, Graph convolutional network}, note = {Published online: 12 March 2026} }
Large-Scale Graphon Learning with Topology-Aware Stochastic Optimization

Seydina Ousmane Niang, Charles Bouveyron, Marco Corneli, and 1 more author

May 2026

preprint

Abs HAL Bib HTML PDF

We introduce SGVAE, a scalable variational graph autoencoder for large-scale network analysis. The decoder is modelled as a neural implicit representation mapping pairs of latent positions to edge probabilities, acting as a continuous graphon estimator. We propose an unbiased stochastic mini-batch ELBO estimator with topology-aware sampling, significantly outperforming state-of-the-art methods in both scalability and clustering quality. A key finding is that the learned latent geometry naturally supports node clustering without additional training, suggesting that decoder architecture is more determinant than encoder sophistication.
@unpublished{niang2026sgvae, title = {Large-Scale Graphon Learning with Topology-Aware Stochastic Optimization}, author = {Niang, Seydina Ousmane and Bouveyron, Charles and Corneli, Marco and Latouche, Pierre}, year = {2026}, month = may, keywords = {Graphon, Variational autoencoder, Large-scale networks, Stochastic optimization, Node clustering, Graph neural networks}, note = {preprint} }
The Deep Zero-Inflated Latent Position Block Model for the Clustering of Nodes in Graphs

Seydina Ousmane Niang, Charles Bouveyron, Marco Corneli, and 1 more author

In Communication dans un Congrès, 2025

Abs HAL Bib HTML

We introduce a deep generative model for the clustering of nodes in sparse networks, combining zero-inflated modelling with a latent position block structure. The model provides simultaneous node clustering and latent space visualisation while accounting for the excess of zeros in network data.
@inproceedings{niang2025deepzilpbm, title = {The Deep Zero-Inflated Latent Position Block Model for the Clustering of Nodes in Graphs}, author = {Niang, Seydina Ousmane and Bouveyron, Charles and Corneli, Marco and Latouche, Pierre}, booktitle = {Communication dans un Congr{\`e}s}, year = {2025}, keywords = {Zero-inflated model, Network clustering, Deep generative model, Latent position model}, }
Importance Weighted Directed Graph Variational Auto-Encoder for Block Modelling of Complex Networks

Seydina Ousmane Niang, Charles Bouveyron, Marco Corneli, and 1 more author

Journal of Statistical Analysis and Data Mining, 2025

Published online: 28 September 2026

Abs HAL Bib HTML PDF

We introduce the first importance weighted graph variational autoencoder for weighted and directed networks: the Deep Zero-Inflated Latent Position Block Model (Deep-ZLPBM). By optimising the importance-weighted ELBO (iw-ELBO) with multiple samples from the variational distribution, we obtain tighter lower bounds, more accurate posterior approximations, and better latent embeddings for clustering and visualisation of complex networks.
@article{niang2025iwgvae, title = {Importance Weighted Directed Graph Variational Auto-Encoder for Block Modelling of Complex Networks}, author = {Niang, Seydina Ousmane and Bouveyron, Charles and Corneli, Marco and Latouche, Pierre}, journal = {Journal of Statistical Analysis and Data Mining}, year = {2025}, keywords = {Importance weighting, Variational autoencoder, Graph neural network, Network clustering, Zero-inflated model}, note = {Published online: 28 September 2026} }