The project ACADEMICS (mAChine LeArning & Data sciEnce for coMplex and dynamICal modelS) aimed to combine Machine Learning (ML) and Data Science (DS) to advance scientific research in two key directions:

• Computing and Information Processing: Developing new theoretical frameworks and learning algorithms tailored for complex scientific challenges involving heterogeneous, irregular, error-prone, dynamic, and intricate datasets, while incorporating relevant prior knowledge.
• Learning Complex and Dynamic Models: Harnessing the synergy between ML and DS to create data-driven models in two scientific domains: climate modeling and the quantitative understanding of social systems. These case studies addressed the critical challenge of learning sophisticated models from abundant, heterogeneous, and dynamic data.

Research Program

The research program focused on achieving meaningful results within the 3-year project duration. The two case studies framed how ML and DS could be integrated to develop relevant models. The methodological challenges addressed included:

• Representation and Model Learning for Complex Data: Identifying sparse latent spaces for complex data or graphs and learning compressed models. Developing methods to detect exceptional phenomena.
• Estimation and Learning from Multi-Source and/or Dynamic Data: Transferring models learned from a source dataset to a related but different target dataset and learning from multi-source complex data.
• Distributed and Adaptive ML for Graphs and Complex Models: Designing distributed, optimization-based learning techniques and adaptive, distributed model inference in high-dimensional spaces.

In climate modeling, the primary question was how to learn effective dynamic models, first using nonparametric ML tools, and second, by combining data from observations and simulations. In computational social sciences, the project explored how to integrate individual features with the structure and dynamics of social networks for better inference of latent correlations, identification of behavioral mechanisms, and modeling of emergent social phenomena.

Project Consortium

The consortium comprised four teams from Université de Lyon (UdL) laboratories, with complementary expertise in ML (deep learning, statistical learning, data mining), DS (complex data analysis, adaptive and data-driven methods, network science), climate modeling, and computational social sciences. The consortium included:

• Laboratoire de Physique (LP): P. Borgnat (Coordinator) and F. Bouchet (PI for Climate).
• Laboratoire Hubert Curien (LabHC): M. Sebban (PI).
• Laboratoire d'Informatique en Images et Systèmes d'information (LIRIS): C. Robardet (PI).
• Laboratoire Informatique du Parallélisme (LIP): P. Gonçalves (PI) and M. Karsai (PI for Computational Social Sciences).

In total, the consortium involved 20 investigators, including the lead scientists. Project Actions

The project’s key actions included:

1. Achieving the scientific research program and disseminating its results through publications and conferences.
2. Enhancing UdL’s international visibility in ML research, a driving force behind the renewal and successes of Artificial Intelligence.
3. Supporting academic training in ML and DS at the Master and Ph.D. levels at UdL.

Impacts and Future Applications

The project’s impacts included developing efficient methods for learning from complex and dynamic data, with applications across multiple domains, such as:

• Technological networks,
• Social sciences and the study of social interactions,
• Climate and environmental sciences,
• Neuroscience and the study of brain networks.

These efforts pushed beyond classical ML applications (time series, 2D images, or text) to achieve a significant scientific breakthrough in learning dynamic models of complex systems.