Towards characterizing the value of edge embeddings in Graph Neural Networks
Authors: Dhruv Rohatgi, Tanya Marwah, Zachary Chase Lipton, Jianfeng Lu, Ankur Moitra, Andrej Risteski
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we show architectures that maintain edge embeddings almost always improve on their node-based counterparts frequently significantly so in topologies that have hub nodes. ... In this section we demonstrate that the representational advantages the theory suggests are borne out by experimental evaluations, both on real-life benchmarks and two natural synthetic tasks we provide. Note that all the experiments were done on a machine with 8 Nvidia A6000 GPUs. |
| Researcher Affiliation | Academia | 1Department of EECS, Massachusetts Institute of Technology, Cambridge, MA, USA 2Department of Machine Learning, Carnegie Mellon University, Pittsburgh, PA, USA 3Department of Mathematics, Duke University, Durham, NC, USA. Correspondence to: Dhruv Rohatgi <EMAIL>, Andrej Risteski <EMAIL>. |
| Pseudocode | No | The paper describes protocols and defines functions mathematically (e.g., Definition 1, Definition 2, Eq. 1, Eq. 2) but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code, nor does it provide a link to a code repository. It refers to existing GNNs and a benchmarking paper for training configurations, but not their specific implementation code. |
| Open Datasets | Yes | We consider several popular GNN benchmarks. Equalizing for all other aspects of the architecture (e.g., depth, dimensionality of the embeddings), we find that the accuracy achieved by edge-based architectures is always comparable and typically slightly better than that of their node-based counterparts. ... Details are included in Section 8.1. ... We consider two synthetic settings to stress test the performance of edge-based architectures. ... we consider a star graph ... Finally, again inspired by the theoretical setting in Theorem 1, we consider probabilistic inference on tree graphs precisely, learning a GNN that calculates node marginals for an Ising model... |
| Dataset Splits | No | The paper states 'For each benchmark, we follow the best performing training configuration as delineated in (Dwivedi et al., 2023).' without explicitly detailing the dataset splits within the main text. |
| Hardware Specification | Yes | Note that all the experiments were done on a machine with 8 Nvidia A6000 GPUs. |
| Software Dependencies | No | The paper mentions using 'Graph Convolutional Network' (GCN) architectures and refers to 'Dwivedi et al., 2023' for training configurations but does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | In Table 2, we show the performance of edge-based and node-based architectures on this dataset, varying the number of leaves n in the star graph and the depth k of the planted edge-based model. In each case, the numbers indicate RMSE of the best-performing edge-based and node-based architecture, sweeping over depths up to 10 (2 the planted model), widths {16, 32, 64}, and a range of learning rates. ... The results we report for each architecture are the best over a sweep of depth {5, 10, 15, 20, 25, 30} and width {10, 32, 64}. |