G-NAC: Graph Neural Automata Clustering via Emergent Domain Formation
G-NAC treats data points as graph cells whose learned local interactions form clusters.
The paper presents an unsupervised method where a recurrent graph-neural cellular rule evolves latent domain states on a fixed neighborhood graph. Those states are turned into a rank-based spectral affinity for partitioning. Across 73 clustering tasks from 57 benchmarks, it reports a mean ARI of 0.7951, roughly matching Genie’s 0.7941 and beating the other tested baselines. Scaling tests from 5,000 to 100,000 nodes were approximately linear in training time and GPU memory, with transfer shown from smaller graphs to matched 100,000-node samples. ArXiv · AI/CL/LG's note
The paper presents an unsupervised method where a recurrent graph-neural cellular rule evolves latent domain states on a fixed neighborhood graph. Those states are turned into a rank-based spectral affinity for partitioning. Across 73 clustering tasks from 57 benchmarks, it reports a mean ARI of 0.7951, roughly matching Genie’s 0.7941 and beating the other tested baselines. Scaling tests from 5,000 to 100,000 nodes were approximately linear in training time and GPU memory, with transfer shown from smaller graphs to matched 100,000-node samples. ArXiv · AI/CL/LG's note
score 5