Table 2.
Results on heterophilic graphs.
| Texas | Wisconsin | Actor | Squirrel | Chameleon | Cornell | |
|---|---|---|---|---|---|---|
| Hom level | 0.11 | 0.21 | 0.22 | 0.22 | 0.23 | 0.30 |
| #Nodes | 183 | 251 | 7,600 | 5,201 | 2,277 | 183 |
| #Edges | 295 | 466 | 26,752 | 198,493 | 31,421 | 280 |
| #Classes | 5 | 5 | 5 | 5 | 5 | 5 |
| GCN† | 55.14 ± 5.16 | 51.76 ± 3.06 | 27.32 ± 1.10 | 31.52 ± 0.71 | 38.44 ± 1.92 | 60.54 ± 5.30 |
| HH-GCN | 71.89 ± 3.46 | 79.80 ± 4.30 | 35.12 ± 1.06 | 47.19 ± 1.21 | 60.24 ± 1.93 | 63.24 ± 5.43 |
| Δ | +16.75 (↑) | +19.04 (↑) | +7.80 (↑) | +15.67 (↑) | +21.80 (↑) | +2.70 (↑) |
| GAT† | 52.16 ± 6.63 | 49.41 ± 4.09 | 27.44 ± 0.89 | 36.77 ± 1.68 | 48.36 ± 1.58 | 61.89 ± 5.05 |
| HH-GAT | 80.54 ± 4.80 | 83.53 ± 3.84 | 36.70 ± 0.92 | 46.35 ± 1.86 | 61.12 ± 1.83 | 72.70 ± 4.26 |
| Δ | +28.38 (↑) | +34.12 (↑) | +9.26 (↑) | +9.58 (↑) | +12.75 (↑) | +10.81 (↑) |
| GraphSAGE† | 82.43 ± 6.14 | 81.18 ± 5.56 | 34.23 ± 0.99 | 41.61 ± 0.74 | 58.73 ± 1.68 | 75.95 ± 5.01 |
| HH-GraphSAGE | 85.95 ± 6.42 | 85.88 ± 3.99 | 36.82 ± 0.77 | 45.25 ± 1.52 | 62.98 ± 3.35 | 74.60 ± 6.06 |
| Δ | +3.51 (↑) | +4.70 (↑) | +2.59 (↑) | +3.64 (↑) | +4.25 (↑) | −1.35 (↑) |
| MixHop† | 77.84 ± 7.73 | 75.88 ± 4.90 | 32.22 ± 2.34 | 43.80 ± 1.48 | 60.50 ± 2.53 | 73.51 ± 6.34 |
| GGCN† | 84.86 ± 4.55 | 86.86 ± 3.29 | 37.54 ± 1.56 | 55.17 ± 1.58 | 71.14 ± 1.84 | 85.68 ± 6.63 |
| H2GCN† | 84.86 ± 7.23 | 87.65 ± 4.98 | 35.70 ± 1.00 | 36.48 ± 1.86 | 60.11 ± 2.15 | 82.70 ± 5.28 |
| MLP† | 80.81 ± 4.75 | 85.29 ± 3.31 | 36.63 ± 0.70 | 28.77 ± 1.56 | 46.21 ± 2.99 | 81.89 ± 6.40 |
We report the test accuracy across many heterophilic graph benchmark datasets, and highlight the absolute improvement in classification accuracy when the model is augmented with Half-Hop.
The “†” results are obtained from (Yan et al., 2021).