Feature-Attention Graph Convolutional Networks for Noise Resilient Learning.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Analysis

Background:

• Real-world information networks often contain noise and inconsistencies.
• Existing graph convolutional networks (GCNs) assume error-free networks and treat node features independently.
• Noisy and sparse node content presents significant challenges for current network analysis methods.

Purpose of the Study:

• To propose a novel framework, Feature-Attention Graph Convolutional Network (FA-GCN), designed to address noisy and sparse node content in networks.
• To enhance feature learning in graph convolutional networks by incorporating attention mechanisms for varying feature importance.
• To improve the robustness and performance of network analysis in the presence of data imperfections.

Main Methods:

• Utilized a Long Short-Term Memory (LSTM) network to generate dense representations of individual node features.
• Introduced a feature-attention mechanism to dynamically weigh the importance of neighboring node features during interaction modeling.
• Employed a spectral-based graph convolution aggregation process to focus on task-relevant neighborhood features.

Main Results:

• FA-GCN demonstrated superior performance compared to state-of-the-art methods in both noise-free and noisy network environments.
• The framework effectively handles networks with varying levels of noise and feature sparsity.
• Experimental validation confirmed the robustness and efficacy of the proposed FA-GCN approach.

Conclusions:

• FA-GCN offers a robust solution for analyzing information networks with noisy and sparse node content.
• The integration of LSTM for feature representation and attention mechanisms for feature importance is key to its success.
• This approach advances the capability of graph convolutional networks for real-world applications where data quality is a concern.