Research map / Graph learning / GNN architectures

Beyond Random Masking: When Dropout meets Graph Convolutional Networks

ICLR 2025 · Graph learning › GNN architectures · approach: Deep graph convolution and simplification

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Theoretical analysis of dropout in graph convolutional networks, arguing its main role is preventing oversmoothing rather than co-adaptation.

Problem

The behavior of dropout in graph convolutional networks is poorly understood, and it is unclear how its role differs from standard neural networks.

Approach

Provides a theoretical analysis showing that dropout in GCNs primarily prevents oversmoothing rather than co-adaptation as in standard networks.

Key results

• Theoretical analysis indicates dropout's primary role in GCNs is preventing oversmoothing rather than co-adaptation, differing from standard neural networks.

Surprising findings

• Dropout's role in GCNs differs fundamentally from its role in standard neural networks (oversmoothing prevention instead of co-adaptation prevention).

Experimental setup

Builds on: GCN dropout

Closest papers by meaning

• Effects of Dropout on Performance in Long-range Graph Learning Tasks (NeurIPS 2025 · similarity 0.91)
• On the Regularization Properties of Structured Dropout (CVPR 2020 · similarity 0.91)
• DropEdge: Towards Deep Graph Convolutional Networks on Node Classification (ICLR 2020 · similarity 0.90)
• DropGNN: Random Dropouts Increase the Expressiveness of Graph Neural Networks (NeurIPS 2021 · similarity 0.90)
• On the Stability of Graph Convolutional Neural Networks: A Probabilistic Perspective (NeurIPS 2025 · similarity 0.88)
• Feature Distribution on Graph Topology Mediates the Effect of Graph Convolution: Homophily Perspective (ICML 2024 · similarity 0.88)