Offre en lien avec l’Action/le Réseau : – — –/Doctorants
Laboratoire/Entreprise : LITIS
Durée : 36 months
Contact : phd-gan_graphs@litislab.fr
Date limite de publication : 2020-09-01
Contexte :
## Position Details
* Location The research will be conducted at LITIS Laboratory (Rouen, France) in Normandy.
The LITIS (EA 4108) is affiliated to Normandie University, University of Rouen and INSA Rouen
Normandie, and founding member of the CNRS Research Federation NormaSTIC.
*Supervisors:
– Benoit Gaüzère, LITIS, INSA Rouen http://pagesperso.litislab.fr/∼bgauzere
– Paul Honeine, LITIS, University of Rouen http://honeine.fr/
* Start date September or October 2020 (or earlier)
* Duration 36 months
Sujet :
Graphs are a powerful and versatile data structure useful to encode
many real-world data, such as networks, molecules and
documents. However, their flexibility comes with some drawbacks
including the complexity associated to elementary operations. For
instance, deciding if two graphs are isomorphic (i.e., structurally
equivalent) or computing a distance between two graphs are NP-Complete
problems, and even hard to approximate. Considering this, several
strategies have been proposed to find some workarounds in order to be
able to process graphs and use them in the Machine Learning
pipeline. The simplest strategy is the explicit embedding of graphs to
an Euclidean space [1], at the cost of losing information. To
overcome this drawback, two major strategies have been recently
investigated. The first one is improving the embedding through kernels
defined on graphs [2,3]. The second and more recent strategy is the
definition of Graph Neural Networks (GNNs) operating directly on
graphs [4–8]. Using neural networks on graphs allows to learn a
proper embedding of graphs given a problem to solve, and then
alleviate the drawbacks of defining a priori an ad hoc embedding.
These embedding-based methods have been commonly investigated for
supervised learning tasks, essentially classification and
regression. However, their interpretability is one major
drawback. Moreover, they were not efficient in many unsupervised
learning tasks, such as estimating the data centroid in k-means or
more generally generating a graph prototype (one graph representative
of a set of graphs, e.g. a median graph). The main reason is the curse
of the preimage, since one needs to reconstruct the solution in the
graph-data space. The preimage problem has already been addressed in
various domains, mainly for kernel-based methods [9,10]. However,
solving this problem for structured data remains an open problem and
only very few attempts have been made on strings and some particular
class of graphs [11]. The purpose of this PhD thesis is to alleviate
the bottlenecks associated to the preimage problem on graphs, through
the use of Generative Adversarial Network (GAN) [12, 13]. GANs
consists of two parts. First, the encoder aims to embed graphs to an
Euclidean space of a predefined dimension. This can be implemented
using existing GNNs and kernel-based methods. Second, the decoder
part aims to reconstruct a graph given a vectorial representation. It
may be considered as the inverse function of the encoder part. The
purpose here is to define this decoder part to take Euclidean spaces
as input and graphs as output, i.e., structured data. By investigating
this approach, the PhD candidate will study particularly molecular
generation.
Profil du candidat :
Required skills
• Master in Applied Mathematics, Computer Science, Data Science, or equivalent
• Experience in Python programming
• Skills in graph theory, neural networks or graph-based methods constitute an advantage
Formation et compétences requises :
Required skills
• Master in Applied Mathematics, Computer Science, Data Science, or equivalent
• Experience in Python programming
• Skills in graph theory, neural networks or graph-based methods constitute an advantage
Adresse d’emploi :
LITIS,
avenue de l’université,
76800 St Etienne du Rouvray
Document attaché : 202004071235_offre.pdf