Graph and Simplicial Complex Prediction Gaussian Process via the Hodgelet Representations
Organizations: University College London · California Institute of Technology · University of Oxford · University of Fribourg ·
Abstract
Predicting labels for graph-structured data is crucial in many scientific applications. Recently, Gaussian processes with graph-level inputs have been proposed as flexible non-parametric models for classification tasks. In this work, we extend this framework to regression tasks and to simplicial complexes, enabling the modelling of attributes defined on edges and higher-order simplices. Drawing on Hodge theory, we construct representations based on the Hodge decomposition, allowing the model to distinguish between exact, co-exact, and harmonic signal components. We introduce Hodgelet representations as rich, learnable topological descriptors of simplicial complexes and show that our framework improves predictive performance across a range of applications. Our work broadens the applicability of Gaussian processes to prediction tasks at the graph and simplicial-complex levels.
Figures & tables
| Accuracy ( ) | Mean squared error ( ) | |||
| MUTAG | AIDS ( ) | FreeSolv | ESOL | |
| GCN (vertex) | ||||
| GAT (hybrid) | \color[rgb]{0.06,0.46,1}\mathbf{0.45}\pm 0.01 | |||
| GIN (hybrid) | ||||
| Transformer (hybrid) | ||||
| WTGP (hybrid) | \color[rgb]{0.06,0.46,1}\mathbf{0.863}\pm 0.010 | \color[rgb]{0.06,0.46,1}\mathbf{99.61}\pm 0.04 | \color[rgb]{0.72,0,0}\mathbf{1.20}\pm 0.04 | \color[rgb]{0.72,0,0}\mathbf{0.41}\pm 0.01 |
| Accuracy ( ) | Mean squared error ( ) | ||
| ieee24-binary | ieee24-multiclass | ieee24-regression ( ) | |
| GCN (vertex) | |||
| GAT (hybrid) | |||
| GIN (hybrid) | \color[rgb]{0.06,0.46,1}\mathbf{0.885}\pm 0.014 | ||
| Transformer (hybrid) | |||
| WTGP (hybrid) | \color[rgb]{0.06,0.46,1}\mathbf{2.19}\pm 0.10 | ||
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Task | Complexes | Vertices | Edges | Triangles |
|---|---|---|---|---|---|
| Div-curl-free | Binary | 100 / 200 | 132.0 | 357.0 | 226.0 |
| Vortices | Binary | 100 / 200 | 132.0 | 357.0 | 226.0 |
| MUTAG | Binary | 188 | 17.9 | 19.8 | 0.0 |
| AIDS | Binary | 2000 | 15.7 | 16.2 | 0.0 |
| FreeSolv | Regression | 628 | 8.9 | 8.6 | 0.0 |
| ESOL | Regression | 1121 | 13.4 | 13.8 | 0.0 |
| MUTAG | AIDS | FreeSolv | ESOL | |
| Hidden dim. | ||||
| Num. layers | ||||
| Dropout | ||||
| Weight decay | ||||
| Readout | Sum | Sum | Mean | Sum |
| Learning rate |
| ieee24-binary | ieee24-multiclass | ieee24-regression | |
| Hidden dim. | |||
| Num. layers | |||
| Dropout | |||
| Weight decay | |||
| Readout | Sum | Sum | Mean |
| Learning rate |
| Method | Model | Training |
|---|---|---|
| Transformer | Architecture described above, hidden dimension , three layers, no dropout, mean readout | Adam, learning rate , weight decay , batch size , at most epochs |
| SNN | Two layers of width with ReLU, polynomial filters of order in ( on MANTRA), scaled by its largest eigenvalue, sum readout, linear head | Adam, learning rate , weight decay , full batch, at most epochs |
| SCNN | As SNN, with an identity term and separate filters of order in and (on MANTRA, in only, since ), each scaled by its largest eigenvalue | As SNN |
| WL-SVM | Weisfeiler–Lehman subtree features ( iterations, label counts normalised to unit length), linear kernel, SVM with and Platt-scaled probabilities | – |
| WL-GP | Features of WL-SVM, scaled linear kernel, constant mean, Bernoulli likelihood, inducing points initialised by -means, KL term not down-weighted ( ) | Adam, learning rate , full-batch epochs, no early stopping |
| MUTAG | AIDS | |||
| NLPD | ECE | NLPD | ECE | |
| GCN (vertex) | \color[rgb]{0.55,0.14,1}\mathbf{0.0229}\pm 0.0009 | |||
| GAT (hybrid) | \color[rgb]{0.72,0,0}\mathbf{0.0226}\pm 0.0007 | |||
| GIN (hybrid) | \color[rgb]{0.06,0.46,1}\mathbf{0.0235}\pm 0.0009 | |||
| Transformer (hybrid) | \color[rgb]{0.06,0.46,1}\mathbf{0.0235}\pm 0.0007 | |||
| WTGP (hybrid) | \color[rgb]{0.55,0.14,1}\mathbf{0.0031}\pm 0.0005 | |||
| ieee24-binary | ieee24-multiclass | |||
| NLPD | ECE | NLPD | ECE | |
| GCN (vertex) | ||||
| GAT (hybrid) | \color[rgb]{0.06,0.46,1}\mathbf{0.031}\pm 0.004 | |||
| GIN (hybrid) | \color[rgb]{0.55,0.14,1}\mathbf{0.028}\pm 0.003 | |||
| Transformer (hybrid) | ||||
| WTGP (hybrid) | \color[rgb]{0.06,0.46,1}\mathbf{0.284}\pm 0.021 | \color[rgb]{0.55,0.14,1}\mathbf{0.024}\pm 0.007 | \color[rgb]{0.55,0.14,1}\mathbf{0.267}\pm 0.008 | \color[rgb]{0.06,0.46,1}\mathbf{0.029}\pm 0.003 |
| FreeSolv | ESOL | ieee24-regression | ||||
| NLPD | ECE | NLPD | ECE | NLPD | ECE | |
| GCN (vertex) | \color[rgb]{0.55,0.14,1}\mathbf{0.030}\pm 0.004 | \color[rgb]{0.55,0.14,1}\mathbf{0.015}\pm 0.001 | \color[rgb]{0.55,0.14,1}\mathbf{0.047}\pm 0.001 | |||
| GAT (hybrid) | \color[rgb]{0.72,0,0}\mathbf{0.041}\pm 0.003 | |||||
| GIN (hybrid) | \color[rgb]{0.55,0.14,1}\mathbf{0.015}\pm 0.003 | |||||
| Transformer (hybrid) | \color[rgb]{0.06,0.46,1}\mathbf{0.018}\pm 0.002 | \color[rgb]{0.06,0.46,1}\mathbf{0.058}\pm 0.008 | ||||
| WTGP (hybrid) | \color[rgb]{0.72,0,0}\mathbf{1.318}\pm 0.023 | \color[rgb]{0.72,0,0}\mathbf{0.939}\pm 0.010 | \color[rgb]{0.06,0.46,1}\mathbf{-2.858}\pm 0.027 | |||
| Dataset | Model | Energy | Sum | Mixture |
|---|---|---|---|---|
| Div-curl-free | HTGP (edge) | |||
| Div-curl-free | WTGP (edge) | |||
| Vortices | HTGP (edge) | |||
| Vortices | WTGP (edge) | |||
| ieee24-binary | HTGP (edge) | |||
| ieee24-binary | WTGP (edge) |
| Dataset | Harmonic | Co-exact | Exact | All |
|---|---|---|---|---|
| Div-curl-free | – | |||
| Vortices | – | |||
| ieee24-binary |
| Dataset | Method | NLPD ( ) | ECE ( ) | Accuracy ( ) |
|---|---|---|---|---|
| Div-curl-free | HTGP (edge) | \color[rgb]{0.72,0,0}\mathbf{0.245}\pm 0.008 | \color[rgb]{0.55,0.14,1}\mathbf{0.041}\pm 0.006 | \color[rgb]{0.72,0,0}\mathbf{0.898}\pm 0.005 |
| Div-curl-free | WTGP (edge) | \color[rgb]{0.55,0.14,1}\mathbf{0.399}\pm 0.038 | \color[rgb]{0.55,0.14,1}\mathbf{0.823}\pm 0.019 | |
| Div-curl-free | SNN | \color[rgb]{0.06,0.46,1}\mathbf{0.516}\pm 0.013 | ||
| Div-curl-free | SCNN | |||
| Div-curl-free | WL-SVM | \color[rgb]{0.06,0.46,1}\mathbf{0.045}\pm 0.013 | ||
| Div-curl-free | WL-GP | \color[rgb]{0.06,0.46,1}\mathbf{0.693}\pm 0.001 | \color[rgb]{0.72,0,0}\mathbf{0.025}\pm 0.010 |
| Dataset | HTGP (edge) with -simplices | HTGP (edge) without -simplices |
|---|---|---|
| Div-curl-free | ||
| Vortices | ||
| ieee24-binary |
| HTGP acc. | Transf. acc. | HTGP NLPD | Transf. NLPD | HTGP ECE | Transf. ECE | |
|---|---|---|---|---|---|---|
| 100 | ||||||
| 500 | ||||||
| 1000 | ||||||
| 1500 | ||||||
| 2000 | ||||||
| 2500 |
| Model | NLPD ( ) | ECE ( ) | Accuracy ( ) | |
|---|---|---|---|---|
| 50 | HTGP (edge) | |||
| WTGP (edge) | ||||
| Transformer | ||||
| 100 | HTGP (edge) | |||
| WTGP (edge) | ||||
| Transformer |
| MUTAG | ieee24-binary | |
|---|---|---|
| 1 | ||
| 5 | ||
| 10 | ||
| 50 | ||
| 100 | ||
| 200 | – |
| Dataset | HTGP ms/epoch | Transf. ms/epoch | Eigendecomposition | Total (HTGP / Transf.) |
|---|---|---|---|---|
| ieee24-binary ( , ) | s | s / s | ||
| ieee24-binary ( , ) | s | s / s | ||
| Div-curl-free ( , ) | s | s / s |
| Method | NLPD ( ) | ECE ( ) | Accuracy ( ) |
|---|---|---|---|
| HTGP (triangle) | |||
| WTGP (triangle) | |||
| SNN | |||
| SCNN | |||
| Transformer |
Gaussian Process Regression
Gaussian process regression uses kernel-defined distributions over functions to make predictions and quantify uncertainty from observed data.
See all 121 papers