Graph Scattering beyond Wavelet Shackles
Authors: Christian Koke, Gitta Kutyniok
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Theoretical results are complemented by numerical investigations: Suitably chosen scattering networks conforming to the developed theory perform better than traditional graph-wavelet based scattering approaches in social network graph classification tasks and significantly outperform other graph-based learning approaches to regression of quantum-chemical energies on QM7. |
| Researcher Affiliation | Academia | Christian Koke Technical University of Munich & Ludwig Maximilian University Munich christian.koke@tum.de; Gitta Kutyniok Ludwig Maximilian University Munich & University of Tromsø kutyniok@math.lmu.de |
| Pseudocode | No | The paper describes its generalized scattering transform iteratively but does not include any structured pseudocode or algorithm blocks with formal labels. |
| Open Source Code | Yes | Yes; please see the supplementary material. |
| Open Datasets | Yes | To aid visual clarity when comparing results, we colour-code the best-performing method in green, the second-best performing in yellow and the third-best performing method in orange respectively. Social Network Graph Classification: To facilitate contact between our generalized graph scattering networks, and the wider literature, we combine a network conforming to our general theory namely Architecture I in Fig. 2 (as discussed in Section 3 with depth N 4, identity as connecting operators and | |-non-linearities) with the low pass aggregation scheme of Section 5 and a Euclidean support vector machine with RBF-kernel (GGSN+EK). The choice N 4 was made to keep computation-time palatable, while aggregation scheme and non-linearities were chosen to facilitate comparison with standard wavelet-scattering approaches. For this hybrid architecture (GGSN+EK), classification accuracies under the standard choice of 10-fold cross validation on five common social network graph datasets are compared with performances of popular graph kernel approaches, leading deep learning methods as well as geometric wavelet scattering (GS-SVM) [12]. More details are provided in Appendix K. |
| Dataset Splits | Yes | For this hybrid architecture (GGSN+EK), classification accuracies under the standard choice of 10-fold cross validation on five common social network graph datasets are compared with performances of popular graph kernel approaches, leading deep learning methods as well as geometric wavelet scattering (GS-SVM) [12]. More details are provided in Appendix K. ... trained with ten-fold cross validation on node and (depending on the model) edge level information. |
| Hardware Specification | Yes | All experiments were run on a single machine with a Ryzen 7 3700x processor, 64 GB of RAM, and a GeForce RTX 2080 Ti GPU. |
| Software Dependencies | No | The paper mentions the use of certain software components (e.g., in Appendix K about using code from [12]), but it does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | In both cases the utilized shift-operator is L : L{λmaxp Lq, node weights satisfy µi 1, the branching ratio in each layer is chosen as 4 and the depth is set to N 4 as well. The connecting operators are set to the identity and non-linearities are set to the modulus (| |). The two architectures differ in the utilized filters, which are repeated in each layer and depicted in Fig. 2. ... Our normal operator is then chosen as L L{λmaxp Lq again. Connecting operators are set to the identity, while non-linearities are fixed to ρně1p q | |. Filters are chosen as psinpπ{2 L q, cospπ{2 L q, sinpπ L q, cospπ L qq acting through matrix multiplication. Output generating functions are set to the identity and depth is N 4 |