Fixed Aggregation Features Can Rival GNNs
Researchers have introduced Fixed Aggregation Features (FAFs), a novel training-free method that reframes graph learning tasks as tabular problems. This approach allows for the use of standard tabular machine learning techniques, enhancing interpretability and flexibility. In evaluations across 14 benchmarks, FAFs combined with tuned multilayer perceptrons matched or surpassed state-of-the-art graph neural networks and transformers on 12 tasks, often utilizing simple mean aggregation. AI
IMPACT Challenges the necessity of complex GNN architectures, suggesting simpler tabular methods can achieve comparable or superior performance on graph data.