Graph Neural Networks with Almost No Features
Explore a groundbreaking approach to handling missing features in graph neural networks presented in this 56-minute conference talk from the Toronto Machine Learning Series. Delve into a simple yet powerful method of feature propagation compatible with any GNN model, outperforming existing approaches in node-classification and link-prediction tasks. Discover how this innovative technique maintains high performance even when 99% of features are missing and scales efficiently to datasets with millions of nodes. Gain insights into the theoretical analysis using compressed sensing tools, understanding how the method acts as a low pass filter and the guarantees for feature reconstruction. Learn from Emanuele Rossi, a Machine Learning Researcher at Twitter and PhD student at Imperial College London, as he shares his expertise in Graph Neural Networks and presents this scalable solution for handling sparse feature sets in graph-based machine learning tasks.