Generalizing CNNs to graphs with learnable neighborhood quantization
Authors: Isaac Osafo Nkansah, Neil Gallagher, Ruchi Sandilya, Conor Liston, Logan Grosenick
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Integrating this approach into a residual network architecture, we demonstrate performance that matches or exceeds other stateof-the-art GNNs on benchmark graph datasets and for predicting properties of nonlinear dynamics on a new finite element graph dataset. |
| Researcher Affiliation | Academia | Isaac Osafo Nkansah1 Neil Gallagher1 Ruchi Sandilya1 Conor Liston1 Logan Grosenick1 1Department of Psychiatry and BMRI, Weill Cornell Medicine, Cornell University, New York, NY, USA |
| Pseudocode | Yes | Algorithm 1 Quantization algorithm: pseudo-code Input: G = (V, E, P) and S, where V is the node set of G, E is the adjacency matrix of G, P is the positional descriptors for all nodes in V and S is the list of subkernels and their angular centroids or quantization ranges. Helper Functions: exnb extracts local graph neighborhoods crad computes relative angular distances gski gets subkernel (by index) a node maps to M {} for a V do Na exnb(G, a) for b Na do r crad(a, b) k_index gski(S, r) M.add( {b, k} ) return M |
| Open Source Code | Yes | Please find the code-base for the paper here: https://github.com/Grosenick-Lab-Cornell/ Quant Nets |
| Open Datasets | Yes | We considered three standard 2D image datasets that vary in complexity: MNIST [21], Fashion-MNIST [41], and CIFAR10 [19]. ... Next, we compared QGRN to SGCN on graph datasets that have positional descriptors, including a novel FEM fixed-mesh graph dataset. The graph benchmark datasets: AIDS, Letters (high/low/med) were post-processed to extract out their positional node descriptors into separate positional attributes ... using inductive learning datasets from Benchmark Data Sets for Graph Kernels [14], namely: AIDS, COIL-DEL, Frankenstein, Enzymes, Letter (low/med/high), Mutagenicity, Proteins, Proteins-Full, Mutag and Synthie. ... We tested on multiple types of node classification datasets, including citation networks (like Cora, Pub Med), Wikipedia hyperlinks networks (like such as the Chameleon dataset) and product relations networks (such as Amazon Computers) [33, 35, 42]. |
| Dataset Splits | Yes | In order to establish equivalence between CNN and QGCN while avoiding full-dataset ceiling effects we separately trained models fit at three different sample sizes (yielding different bias-variance trade-offs) by varying the dataset train-test splits (see Appendix F). ... Appendix F (table 8) shows the expanded version of 1, showing different train-test splits, devised to explore bias-variance trade-offs. ... The splits chosen were: (train, test) = (100, 20), (1000, 200) and (10000, 1000), with equal sampling across categories. |
| Hardware Specification | Yes | All experimental results provided in this paper were the results of runs on Google Colab premium offering of GPUs and TPUs. As much as specific details on these systems are publicly available, we have tabulated them below. The two main accelerator options we used were the A100 GPUs and TPUs. In table 26, we outline some of the system specifications. ... Table 26: Compute Resources. Google Colab Pro+ offering. System System RAM (GB) GPU RAM (GB) Disk Space (GB) Nvidia A100 83.5 40.0 201.2 Google T4 GPU 51.0 15.0 201.2 Google L4 GPU 62.8 22.5 201.2 Google TPU 334.6 NA 225.3 |
| Software Dependencies | No | The paper mentions using the "Adam optimizer [16]" and that experiments were implemented in "Py Torch Geometric format" and "FEni CS library" for FEM simulations, but it does not specify version numbers for these software components or any other key libraries for reproducibility. |
| Experiment Setup | Yes | All models were trained 5 times on each dataset, with different random parameter initializations and random ordering of the training data for each run, using cross-entropy loss and the Adam optimizer [16] with a learning rate of 0.01 for 200 epochs. |