Improving Subgraph-GNNs via Edge-Level Ego-Network Encodings
Authors: Nurudin Alvarez-Gonzalez, Andreas Kaltenbrunner, Vicenç Gómez
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In an empirical evaluation on four benchmarks with 10 graph datasets, our results match or improve previous baselines on expressivity, graph classification, graph regression, and proximity tasks while reducing memory usage by 18.1x in certain real-world settings. 7 Experimental Results |
| Researcher Affiliation | Academia | Nurudin Alvarez-Gonzalez EMAIL Universitat Pompeu Fabra Andreas Kaltenbrunner EMAIL Universitat Oberta de Catalunya ISI Foundation Turin Vicenç Gómez EMAIL Universitat Pompeu Fabra |
| Pseudocode | Yes | Algorithm 1 Elene Node Encoding using BFS. |
| Open Source Code | Yes | We describe our experimental protocol in 7.1 and provide reproducible code, hyper-parameters, and analysis scripts through Github6 for four experimental benchmarks: 6https://github.com/nur-ag/ELENE |
| Open Datasets | Yes | A) Expressivity. Evaluates whether models distinguish non-isomorphic graphs (on 1-WL EXP (Abboud et al., 2021) and 3-WL SR25 (Balcilar et al., 2021) equiv. datasets), count sub-graphs (in Random Graph (Chen et al., 2020)), and evaluate graph-level properties (Corso et al., 2020). B) Proximity. Measures whether models learn long-distance attributed node relationships in h-Proximity datasets (Abboud et al., 2022). C) Real World Graphs. Evaluates performance on five large-scale graph classification/regression datasets from Benchmarking GNNs (ZINC, CIFAR10, PATTERN) (Dwivedi et al., 2020), and the Open Graph Benchmark (Mol HIV, Mol PCBA) (Hu et al., 2020a). |
| Dataset Splits | Yes | Table 6: Dataset statistics. ... Nr. of Graphs (Train / Valid / Test) ... ZINC-12K ... 10000 / 1000 / 1000 ... Mol PCBA ... 350343 / 43793 / 43793 |
| Hardware Specification | Yes | Environment. Experiments ran on a shared server with a 48GB Quadro RTX 8000 GPU, 40 CPU cores and 502GB RAM. ... we also reproduced our experiments on real-world graphs on a SLURM cluster with nodes equipped with 22GB Quadro GPUs. Finally, scalability experiments ran on Tesla T4 GPUs with 15.11GB of VRAM... |
| Software Dependencies | No | The paper mentions several GNN architectures (GIN, GINE, GNN-AK, etc.) but does not specify software dependencies like Python, PyTorch, or CUDA with version numbers in the text. |
| Experiment Setup | Yes | We explore sub-sets of Elene hyper-parameters via grid search with k {0, 1, 2, 3, 5} parameter ranges for Elene and Elene-L, and test the ED/ND variants for Elene-L with embedding params. ω {16, 32, 64}, ρ = dmax, using masked-mean pooling for stability. All models use L = 4 for EXP and L = 2 for SR25. Table 5: Hyper-parameters controlling the behaviour of explicit ego-network attribute encodings. |