Understanding Truncated Positional Encodings for Graph Neural Networks
A new research paper explores the theoretical properties of truncated positional encodings (PEs) used in graph neural networks (GNNs). The study reveals that while complete spectral and walk-based PEs offer equivalent expressive power, their truncated versions, commonly used in practice due to complexity constraints, diverge significantly. The research demonstrates that truncated spectral PEs are less powerful than previously thought and that a combination of different truncated PEs performs best on real-world datasets. AI
IMPACT This research clarifies the expressive power of commonly used truncated positional encodings in GNNs, potentially guiding more effective model design.