Tag: gnn
All the talks with the tag "gnn".
The Expressive Power of GNNs
Pankaj KumarPublished: at 02:00 PMThis talk delves into the theoretical foundations of Graph Neural Networks (GNNs), focusing on their ability to model complex relationships between node features and graph structures. While GNNs excel in tasks like node classification and link prediction, their expressive power remains challenging to quantify due to the interplay of graph topology and node features. The discussion highlights key challenges, such as trade-offs between expressiveness and computational efficiency, and explores advancements like the connection between GNNs and the Weisfeiler-Lehman graph isomorphism test. The talk also outlines future directions for developing rigorous evaluation methods to better understand and assess GNN expressivity.
Private Graph Extraction via Feature Explanation
Rishi Raj SahooPublished: at 02:00 PMThis study explores privacy leakage in GNNs through graph reconstruction attacks using feature explanations. It compares explanation methods—gradient-based, perturbation-based, and surrogate model-based—showing that explanations aid graph reconstruction, with a trade-off between privacy and utility. A defense using randomized response is proposed to reduce attack success.
GNNX-BENCH - Perturbation-based GNN Explainers
Rishi Raj SahooPublished: at 02:00 PMThis work presents a benchmark for Graph Neural Network Explainability with perturbation. This involves some metrics like Sufficiency, Necessity, Stability, and Reproducibility to observe which method performs better. Finally, the work provides recommendations for choosing methods for particular graph tasks.
Quaternion Graph Neural Networks
Rucha Bhalchandra JoshiPublished: at 02:00 PMRecently, graph neural networks (GNNs) have become an important and active research direction in deep learning. This talk proposes Quaternion Graph Neural Networks (QGNN) to learn graph representations within the Quaternion space, a hyper-complex vector space. The talk covers state-of-the-art results on benchmark datasets for graph classification and node classification, as well as knowledge graph completion.