Edouard Pineau
About me
I’m a Ph.D. student at Télécom Paris (Institut Polytechnique de Paris), working in the DIG research group under the supervision of Thomas Bonald. My thesis is supported by the company Safran through the CIFRE convention 2017/1317, with the cosupervision of Sébastien Razakarivony from the data analysis team of Safran Research and Technology.
I’m interested in recent advances in representation learning, with particular focus on time series and graphs representation.
Education
2017-2020: PhD Computer Science
Institut Polytechnique de Paris
2016-2017: Msc Machine Learning
Paris-Dauphine University & ENS Paris-Saclay
2014-2016: Msc Applied Mathematics for Financial Engineering
Paris-Dauphine University
2010-2013: Bsc Applied Mathematics and Economics
Paris-Dauphine University
Recent published research
Unsupervised ageing detection on a causality graph
Short abstract: we assume that a dataset of multivariate time series (MTS) has an underlying causal structure that we can exploit to represent samples. Our contribution is a new representation framework that consists of first finding the overall causality graph G in a studied dataset and then mapping each sample onto G to obtain a causality-based representation. Since causality is an important feature underlying MTS data, we claim and show that representating samples on G is meaningful. We apply our model on health monitoring tasks, using two MTS datasets coming from ageing mechanical systems.
Time series source separation with slow flows
Short abstract: we show that slow feature analysis (SFA), a common time series decomposition method, naturally fits into the flow-based models (FBM) framework, a type of invertible neural latent variable models trained by exact maximum likelihood. Building upon recent advances on blind source separation, we show that such a fit makes the time series decomposition identifiable.
Seq2VAR: multivariate time series representation with relational neural networks and linear autoregressive model
Short abstract: linear vector autoregressive models (VAR) are common multivariate time series (MTS) modeling tools. In particular, sparifying VAR models enable to extract information about causality underlying the data. Yet, when we have a large dataset of MTS, fitting a sparse VAR in each sample is not tractable. We propose to leverage de recent advances on time series representation with relational neural network to show that we can robustly and efficiently extract linear causality from MTS samples.
Variational Recurrent Neural Networks for Graph Classification
Short abstract: graphs can be seen as sequences of nodes. With node ordering methods and recurrent neural networks (i.e. neural networks made for sequence modeling), we can greedily learn representation of graphs for classification.
A simple baseline algorithm for graph classification
Short abstract: graphs have exotic features like variable size and no trivial alignement. To compare graphs, they need to be represented with features that are invariant no node ordering and consistent with changes in graphs (e.g. node or edge addition). A feature that satistisfies the aforementioned properties is the Laplacian spectrum. We show that it is a simple but competitive baseline for graph classification.
A deeper analysis of the results is given here.