Publications

Based on connections with the Partial Information Decomposition framework, we introduce a novel GNN layer, namely the Graph Ordering Attention (GOAT) that imposes neighborhood orderings according to the attention coefficients. Read more

We derive a theoretical analysis on the Lipschitz continuity of attention and show that enforcing Lipschitz continuity through normalization can significantly improve the performance of deep attention models. Read more

We propose a parametrised graph shift operator (PGSO) to encode graphs, providing a unified view of common GSOs, and improve GNN performance by including the PGSO into the training in an end-to-end manner. Read more

Moving beyond local interactions, nodes can share structural similarities, based on their position. We investigate feature augmentation methods of graph neural networks using structural entropy measures. Read more

Based on topological criteria and, specifically the separability, we introduce a universal approximation scheme of continuous functions on graphs. It is based on the disambiguation of identical node attributes. Read more

We propose a graph-based initialization of neural networks extending graph degeneracy observations. Such an initialization can encourage neurons that have structural importance in the neural network. Read more

We study how we can approximate graph distances by aligning adjacency matrices in a corpus of graphs. In order to allow the differentiable optimization, we suggest the utilization of soft permutation matrices. Read more

Solving simultaneously link prediction and community detection is important in recommendation systems. Here, we show how we can extend the information that Graph Auto-encoders process towards this direction. Read more

Standard GNNs use a 1-hop aggregation per layer, limiting their ability to capture graph properties. We iteratively extend the aggregation operator of graph neural networks to increase their receptive field. Read more