Special Sessions

Special Session on “Recent Advances in Deep Learning for Graphs”

Special Session Organizers:

Ming Li, Zhejiang Normal University, China. mingli@zjnu.edu.cn

Alessio Micheli, University of Pisa, Italy.  micheli@di.unipi.it

Giorgio Stefano Gnecco, IMT School for Advanced Studies, AXES Research Unit, Lucca, Italy. giorgio.gnecco@imtlucca.it

Marcello Sanguineti, University of Genoa, Italy.  marcello.sanguineti@unige.it

Federico Errica, NEC Laboratories Europe, Germany and U. of Pisa, Italy. federico.errica@neclab.eu

Paper Submission Deadline: Tuesday May 3, 2022 (AoE)

https://easychair.org/my/conference?conf=lod2022

The field of deep learning for graphs studies how to extend deep learning techniques to data that are represented as a graph. A graph is a data structure able to capture intricate forms of interrelations between a set of possibly high-dimensional entities, and as such its use is frequent in domains like biology, physics, computer graphics, chemistry, and social networks, to name a few. This poses two compelling challenges: on the one hand, one wants to design methods that allow us to better study and comprehend the complex available data; on the other hand, we need an efficient and effective way to capture and encode the geometric patterns of the graphs. To address them, different styles of deep neural networks for graphs
(DNNG) have been proposed, spanning from the recursive approach of Graph Neural Networks to the multi-layered Convolutional Neural Networks for Graphs. With the term Deep Graph Networks (DGN), we can denote a broader class of models which includes neural, Bayesian, and generative deep learning models for graphs. In their various incarnations, DGNs have become a topic of immense interest at both theoretical and practical levels, thanks to the remarkable richness of the learned graph representations when applied to the tasks of node/graph classification, graph generation, and link prediction. DGNs and
related methods can be used in countless applications, ranging from molecule properties (e.g., energy or biological activities) prediction and drug repurposing to recommender systems in e-commerce and intelligent transportation. The ever increasing stream of publications about the topic of deep learning for graphs is indicative of the research questions that arise at both academic and industrial levels, where
there is a considerable demand for more advanced technologies and algorithms, theoretical foundations, as well as tools and platforms for real-world applications that can overcome well-known issues in the field. The goal is, therefore, to develop a broad class of new models and algorithms of different nature, be it spectral, spatial, recursive, or a combination of these, which can thrive in highly competitive and practical scenarios such as large-scale, dynamic, and ambiguous graphs with missing node features or edges. Proper mathematical formulations and sound theories are also needed to characterize the power and limitations of DGNs. Put simply, this special session focuses on theoretical studies, model designs, algorithmic developments, and advanced applications of the field of deep learning for graphs.

Topics of Interests
This special session aims to provide a forum for both the academic and industrial
communities to report recent results related to (advanced) deep graph networks from the
perspectives of theory, models, algorithms and applications. Topics appropriate for this
special issue include (but are not necessarily limited to):

• Deep Neural Networks for Graphs and Deep Graph Networks novel approaches
• Novel DNNG Architectures
• Deep Learning for Graphs
• Graph Representation Learning
• Fast and/or Distributed Learning Algorithms for DNNG/DGNs
• Spectral-based and Spatial-based Methods
• Novel Approaches to Graph Convolution, Recursion on Graphs
• Novel Graph Pooling, Graph Attention, Message Passing Mechanisms on Graphs
• Algorithms for Pre-Trained DNNG/DGNs, Graph Transformer Networks
• Spectral Graph Theory, Graph Wavelets
• Expressive, Generalization and Representational Power of DNNG/DGNs
• Universality of Invariant or Equivariant DNNG/DGNs
• Learning Theory on DNNG/DGNs
• Probabilistic and Generative Models of Graphs
• Combination of Neural Networks for Graphs and Gaussian Processes
• Heterogeneous DNNG/DGNs, Hyper-DNNGs, Multi-View DNNGs
• Novel Learning Frameworks for Graph Classification, Node Classification, Link Prediction
• Adversarial Attacks and Defenses on Graphs
• Trustworthy Approaches for Deep Learning on Graphs
• Applications based on Novel DNNG/DGNs to Physics, Mathematics, Chemistry, Biology, Computer Vision, Nature Language Processing, Social Networks, Traffic Networks, Communication Networks, Internet of Things, Recommender Systems, Urban Computing, Drug Design, and their potential cross-modality, etc.