The paper reports that SLIMG, a linear GNN designed for node classification, degrades substantially when applied to link prediction on heterophily graphs. On five real-world heterophily datasets (chameleon, squirrel, actor, twitch, pokec), SLIMG scores range from 12.0 to 76.9 hits@100/1000, while on homophily graphs it reaches up to 86.8. On synthetic off-diagonal (heterophily) structures with local features, SLIMG drops from 82.5 to 31.1 hits@100. The authors attribute this to SLIMG's inability to adjust node similarity for embeddings where connected nodes are expected to be dissimilar.
Evidence
correlational
Key metric
SLIMG hits@100: chameleon 76.9±2.8, squirrel 19.6±1.5, actor 18.7±1.0, twitch 12.0±0.3, pokec 21.7±0.2 (heterophily); synthetic local x off-diag a: 31.1±1.1 vs local x diag a: 82.5±1.6
Caveat
The paper notes that SLIMG was designed for node classification, not link prediction, so the failure may be a task mismatch rather than a fundamental architectural limitation.