Self-Attention in Colors: Another Take on Encoding Graph Structure in Transformers

TL;DR

This paper introduces CSA, a novel self-attention mechanism for graph transformers that encodes structural and topological information without message passing, achieving state-of-the-art results on molecular datasets.

Contribution

The paper presents CSA, a new self-attention method that modulates feature channels and encodes graph structure using random walk-based positional encoding, bypassing message passing.

Findings

Achieves state-of-the-art results on ZINC dataset.

Effectively encodes higher-order topological features.

Provides a flexible framework for graph structure encoding.

Abstract

We introduce a novel self-attention mechanism, which we call CSA (Chromatic Self-Attention), which extends the notion of attention scores to attention _filters_, independently modulating the feature channels. We showcase CSA in a fully-attentional graph Transformer CGT (Chromatic Graph Transformer) which integrates both graph structural information and edge features, completely bypassing the need for local message-passing components. Our method flexibly encodes graph structure through node-node interactions, by enriching the original edge features with a relative positional encoding scheme. We propose a new scheme based on random walks that encodes both structural and positional information, and show how to incorporate higher-order topological information, such as rings in molecular graphs. Our approach achieves state-of-the-art results on the ZINC benchmark dataset, while providing a flexible framework for encoding graph structure and incorporating higher-order topology.