Learning Graph-Conditioned Predictors in Weight Space
Organizations: Australian National University, Australia · CSIRO Technology, Australia
Abstract
The same node attributes can imply different labels in different structural contexts. Yet fast graphless models typically treat topology as additional evidence supplied to a globally shared predictor. This assumes that structure augments the available information, rather than determining how node attributes should be interpreted. We introduce SCOPE (Structure-COnditioned PrEdictor), in which a small generator maps each node's cached structural fingerprint to coefficients over a shared low-rank basis, producing a context-specific update to a global weight matrix. The base predictor, basis, and generator remain shared across all nodes, while the resulting transformation adapts how attributes are processed in each relational environment. Once fingerprints are cached, inference requires only a forward pass and no online message passing. This design combines context-specific specialization with global parameter sharing, compact low-rank adaptation, and message-passing-free inference. On two large OGB node-classification graphs, conditioning the weights with raw positional signals yields clear gains over supplying the same signals through input concatenation. Further results under a structure-transductive held-out-node protocol and in teacher-free graph classification show that the mechanism extends beyond a particular distillation setting. These findings establish graph structure not only as predictive evidence, but as context that determines the predictor used to interpret that evidence.
Figures & tables
| Variant | Acc. | vs. GLNN |
|---|---|---|
| ( M1 ) structure-free | ||
| GLNN (plain MLP KD) | 58.98 0.53 | 0.00 |
| ( M2 ) structure as input | ||
| + concatenate | 63.46 0.30 | +4.48 |
| ( M3 ) structure as adaptation context | ||
| without conditioning (controls) | ||
| ogbn-arxiv | ogbn-products | ||||||
| Signal (dim) | Used as | tran | ho-full | tran | ho-full | ||
| Teacher (SAGE) | message passing | 71.08 0.42 | 70.52 0.36 | 78.15 0.39 | 78.18 0.26 | ||
| none (GLNN) | – | 64.81 0.28 | 58.98 0.53 | 0.0 | 64.60 0.34 | 64.16 0.19 | 0.0 |
| node-wise spectral structural summaries | |||||||
| HKS (8) | generation | 66.11 0.36 | 59.51 0.63 | +0.5 | 63.70 0.30 | 63.15 0.26 | |
| RWSE (8) | generation | 66.04 0.25 | 59.75 0.44 | +0.8 | 63.68 0.46 | 63.15 0.29 | |
| Config | struct-gain | frac. helped | node-spec. | frac. own better | |
|---|---|---|---|---|---|
| 0.388 | 0.647 | 0.854 | 1.896 | 0.891 | |
| 0.358 | 0.691 | 0.870 | 2.041 | 0.900 | |
| 0.315 | 0.726 | 0.888 | 2.074 | 0.915 | |
| thick | 0.329 | 0.764 | 0.893 | 2.366 | 0.922 |
| Transductive | Held-out nodes (ho-full) | |||||||||||||
| Method | Cora | Citeseer | Pubmed | A-computer | A-photo | ogbn-arxiv | ogbn-products | Cora | Citeseer | Pubmed | A-computer | A-photo | ogbn-arxiv | ogbn-products |
| Graph-based task references | ||||||||||||||
| Teacher (SAGE) | 80.49 1.90 | 70.23 2.15 | 75.42 1.85 | 82.11 1.88 | 90.71 1.26 | 71.08 0.42 | 78.15 0.39 | 81.24 1.91 | 70.00 2.91 | 75.37 2.74 | 82.91 1.56 | 91.04 1.21 | 70.52 0.36 | 78.18 0.26 |
| SGC | 80.81 1.37 | 70.91 2.36 | 74.53 1.91 | 83.74 1.58 | 91.13 1.12 | 69.31 0.31 | 68.26 0.15 | 80.33 2.13 | 70.83 3.02 | 72.94 2.63 | 83.58 1.84 | 91.31 0.81 | 69.07 0.69 | 68.45 0.16 |
| Learned positional-input reference (CE, no teacher) | ||||||||||||||
| DW+MLP | 75.50 1.25 | 62.35 2.80 | 67.98 2.75 | 76.08 1.86 | 79.87 1.51 | 69.45 0.25 | 78.06 0.23 | 71.38 2.78 | 63.92 3.62 | 65.34 2.96 | 80.30 1.43 | 81.13 1.72 | 69.81 0.35 | 78.88 0.43 |
| ogbn-arxiv | ogbn-products | ||||
|---|---|---|---|---|---|
| Method | Structural route | tran | ho-obs | tran | ho-obs |
| TINED ( Zhou et al., 2025 ) | teacher-weight transfer | 64.44 0.72 | 59.79 0.46 | 69.48 0.25 | 68.68 0.27 |
| SCOPE | conditioned weights | 68.87 0.40 | 63.73 0.45 | 70.52 0.30 | 69.44 0.26 |
| NOSMOG ( Tian et al., 2023 ) | DeepWalk input | 71.65 0.29 | 70.09 0.55 | 78.45 0.38 | 77.02 0.19 |
| SCOPE + DeepWalk | DeepWalk-conditioned weights | 71.56 0.30 | 69.98 0.47 | 78.93 0.44 | 78.63 0.34 |
| TINED+ ( Zhou et al., 2025 ) | teacher transfer + DeepWalk input | 71.52 0.34 | 70.42 0.35 | 78.59 0.28 | 79.31 0.29 |
| ROC-AUC | 10-fold accuracy | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Arm | molhiv | PROTEINS | DD | IMDB-B | REDDIT-B | NCI1 | ||||||
| no structural signal | ||||||||||||
| plain MLP | 68.3 | 0.7 | 60.6 | 2.2 | 70.3 | 3.8 | 70.2 | 3.6 | 62.8 | 1.6 | 57.1 | 1.4 |
| structure as input | ||||||||||||
| fingerprint-only | 70.4 | 0.6 | 74.2 | 3.8 | 76.2 | 2.9 | 70.1 | 1.6 | 83.3 | 2.2 | 70.0 | 2.4 |
| input-concat | 76.7 | 0.9 | 68.4 | 4.0 | 72.6 | 3.3 | 71.8 | 4.0 | 78.8 | 2.0 | 68.7 | 2.1 |
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | nodes | edges | feat. dim | classes | teacher | student | lr | wd | drop. | / |
|---|---|---|---|---|---|---|---|---|---|---|
| Cora | 2,485 | 5,069 | 1,433 | 7 | 2 128 | 2 128 | 0.010 | 5e-3 | 0.6 | 8 / 18 |
| Citeseer | 2,110 | 3,668 | 3,703 | 6 | 2 128 | 2 128 | 0.010 | 1e-3 | 0.1 | 8 / 18 |
| Pubmed | 19,717 | 44,324 | 500 | 3 | 2 128 | 2 128 | 0.005 | 1e-3 | 0.4 | 8 / 18 |
| A-computer | 13,381 | 245,778 | 767 | 10 | 2 128 | 2 128 | 0.003 | 5e-3 | 0.1 | 8 / 18 |
| A-photo | 7,487 | 119,043 | 745 | 8 | 2 128 | 2 128 | 0.002 | 1e-3 | 0.3 | 8 / 18 |
| ogbn-arxiv | 169,343 | 1,166,243 | 128 | 40 | 3 256 BN | 3 256 BN | 0.010 | 0 | 0.2 | 32 / 7 |
| Dataset | graphs | avg. nodes | hidden | layers | batch | epochs |
|---|---|---|---|---|---|---|
| ogbg-molhiv | 41,127 | 25.5 | 300 | 3 | 256 | 100 |
| PROTEINS | 1,113 | 39.1 | 128 | 2 | 128 | 100 |
| DD | 1,178 | 284.3 | 128 | 2 | 64 | 100 |
| IMDB-B | 1,000 | 19.8 | 128 | 2 | 128 | 100 |
| REDDIT-B | 2,000 | 429.6 | 128 | 2 | 64 | 100 |
| NCI1 | 4,110 | 29.9 | 128 | 2 | 128 | 100 |
| Method | Cora | Citeseer | Pubmed | A-computer | A-photo |
|---|---|---|---|---|---|
| Transductive | |||||
| TINED ( Zhou et al., 2025 ) | 82.63 1.57 | 74.43 1.53 | 77.09 2.14 | 85.18 1.12 | 93.97 0.53 |
| SCOPE | 80.60 2.16 | 72.11 1.46 | 76.25 1.88 | 82.42 1.89 | 92.06 0.93 |
| NOSMOG ( Tian et al., 2023 ) | 83.04 1.26 | 73.78 1.54 | 77.34 2.36 | 84.04 1.01 | 93.36 0.69 |
| SCOPE + DeepWalk | 80.57 2.45 | 72.18 1.32 | 76.25 1.96 | 82.60 2.02 | 91.23 0.94 |
| TINED+ ( Zhou et al., 2025 ) | 83.70 1.02 | 75.39 1.59 | 77.75 3.14 | 84.82 1.58 | 94.05 0.39 |
| Variant | tran | ho-full |
|---|---|---|
| (a) conditioned: rank sweep on ogbn-products | ||
| SCOPE | 70.10 | 69.64 |
| SCOPE | 70.63 | 70.06 |
| (b) unconditioned: Low-rank control vs. GLNN, both graphs | ||
| Low-rank (arxiv) | 53.60 | 53.45 |
| Low-rank , matched (arxiv) | 60.08 | 58.05 |
| Constraint added (cumulative) | tran | ho-full |
| none ( SCOPE , all learned) | 69.00 | 64.50 |
| teacher | 68.85 | 64.01 |
| + frozen | 65.21 | 63.35 |
| + SVD( ) | 65.54 | 63.45 |
| + frozen (fully constrained) | 61.87 | 60.99 |
| control: SVD basis, orthogonalized | 64.72 | 63.71 |
| Config | 16 | 32 | 64 | 128 | |
|---|---|---|---|---|---|
| 1.41 | 1.10 | 1.48 | 1.15 | 1.09 | |
| 0.33 | 0.52 | 0.51 | 0.45 | 0.46 | |
| 0.23 | 0.24 | 0.21 | 0.21 | 0.21 | |
| thick | 0.67 | 0.70 | 0.60 | 0.52 | 0.56 |
| val | test | |
| (a) main comparison | ||
| GIN teacher, plain (ref) | 80.56 | 73.57 |
| plain MLP (CE) | 72.18 | 68.64 |
| plain MLP (KD) | 72.07 | 66.89 |
| per-graph SCOPE (CE, ) | 79.67 | 71.58 |
| per-graph SCOPE (KD) | 79.45 | 70.90 |
Graph Representation Learning
Graph representation learning develops methods for encoding nodes, edges, and whole graphs for prediction, comparison, and transfer.
See all 320 papers