Previous | 
Home
Section: Highlights of the Year
Highlights of the Year
Machine Learning & Data Science for Complex and Dynamical Models
The Dante team is part of a consortium (including the LIP, the Physics Lab from ENS de Lyon, the LabHC from U. Jean Monnet and LIRIS from U. Lyon 1) that got funded a 4 years project within the call “Scientific Breakthrough” of IDEX de Lyon.
With a total envelope of 1.2M euros, the project "mAChine LeArning & Data sciEnce for coMplex and dynamICal modelS" (ACADEMICS) combines Machine Learning (ML) and Data Science (DS) for the purpose of scientific research into two challenging directions:
-
Computing and information processing – develop new theoretical frameworks and learning algorithms adapted to difficult scientific contexts involving heterogeneous, irregular, error-prone, dynamic and complex data, while taking into account prior knowledge whenever it is relevant.
-
Complex and dynamic models learning – leverage the synergy between ML and DS to devise data-driven models in two scientific domains: climate modeling, and quantitative understanding of social systems. Focusing on these two case studies, the project will tackle the key issue of how to learn intricate models from numerous, heterogeneous and dynamic data.
The research program is elaborated along specific scientific issues that can reasonably lead to significant results within the 3-year lifetime of the project. The two case studies are instrumental to frame the way ML and DS can combine to yield relevant models. The methodological axes are:
-
Representation and model learning for complex data: How to find sparse latent spaces for complex data or graphs, and how to learn compressed models? How to identify exceptional phenomena?
-
Estimation and learning from multi-source and/or dynamic data: How to transfer a model learned from source data to related but different target data? How to learn from multi-source complex data?
-
Distributed and adaptive machine learning for graphs and complex models: How to design distributed optimization-based learning? How to develop adaptive and distributed model inference in high dimension?
In close connexion with these methodological questions, the climate modeling use-case raises the central interrogation of how to learn effective dynamic models, firstly in a nonparametric way by means of ML tools and secondly, by mixing several data sources (from observations and simulations). As for computational social science, the challenge is to embed together in ML approaches, individual features, global structures and dynamics of social networks. The goal here, is to benefit from their complementarity to infer latent correlations, to identify behavioral mechanisms and to better model emergent social phenomena.
Books on Dynamic Networks by Márton Karsai
After a book chapter on Control Strategies of Contagion Processes in Time-varying Networks in Temporal Network Epidemiology in collaboration with Nicola Perra [65], a full book on Bursty Human Dynamics was just released at the end of the year in collaboration with Hang-Hyun Jo and Kimmo Kaski [40].