Any2Graph: Deep End-To-End Supervised Graph Prediction With An Optimal Transport Loss
Authors: Paul Krzakala, Junjie Yang, Rémi Flamary, Florence d'Alché-Buc, Charlotte Laclau, Matthieu Labeau
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments showcase the versatility of the approach that outperforms existing competitors on a novel challenging synthetic dataset and a variety of real-world tasks such as map construction from satellite image (Sat2Graph) or molecule prediction from fingerprint (Fingerprint2Graph). |
| Researcher Affiliation | Academia | Paul Krzakala LTCI & CMAP , Télécom paris, IP Paris Junjie Yang LTCI, Télécom paris, IP Paris Rémi Flamary CMAP, Ecole polytechnique, IP Paris Florence d Alché-Buc LTCI, Télécom paris, IP Paris Charlotte Laclau LTCI, Télécom paris, IP Paris Matthieu Labeau LTCI, Télécom paris, IP Paris |
| Pseudocode | No | The paper describes computational procedures like the conditional gradient algorithm but does not present them in structured pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | 1All code is available at https://github.com/Krzakala Paul/Any2Graph. |
| Open Datasets | Yes | The first one, Coloring, is a new synthetic dataset that we proposed, inspired by the four-color theorem. The input is a noisy image partitioned into regions of colors and the goal is to predict the graph representing the regions as nodes (4 color classes) and their connectivity in the image. An example is provided in Figure 5 and more details are in Appendix D. Then, we consider four real-world benchmarks. Toulouse [5] is Sat2Graph datasets where the goal is to extract the road network from binarized satellite images of a city. USCities is also a Sat2Graph dataset but features larger and more convoluted graphs. Note that we leave aside the more complex RGB version of USCities as it was shown to require complex multi-level attention architecture [36], which is beyond the scope of this paper. Finally, following Ucak et al. [42], we address the Fingerprint2Graph task where the goal is to reconstruct a molecule from its fingerprint representation (list of tokens). We consider two widely different datasets for this tasks: QM9 [50], a scarce dataset of tiny molecules (up to 9 nodes) and GBD13Blum and Reymond [7], a large dataset 2 featuring molecules with up to 13 heavy atoms. |
| Dataset Splits | Yes | Additional details concerning the datasets (e.g. dataset size, number of edges, number of nodes) are provided in Appendix E.1. ... Table 4: Table summarizing the properties of the datasets considered. ... (TRAIN/TEST/VALID) |
| Hardware Specification | Yes | All values are computed on NVIDIA V100/Intel Xeon E5-2660. |
| Software Dependencies | No | The paper mentions using the 'Adam optimizer' and specific neural network architectures, but it does not provide specific version numbers for software libraries, deep learning frameworks (e.g., PyTorch, TensorFlow), or programming languages (e.g., Python). |
| Experiment Setup | Yes | All hyperparameters are given in Table 5. |