Stochastic Iterative Graph Matching
Authors: Linfeng Liu, Michael C Hughes, Soha Hassoun, Liping Liu
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments across synthetic graph datasets as well as biochemistry and computer vision applications. Across all tasks, our results show that SIGMA can produce significantly improved graph matching results compared to state-of-the-art models. Ablation studies verify that each of our components (stochastic training, iterative matching, and dummy nodes) offers noticeable improvement. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Tufts University, MA, USA 2Department of Chemical and Biological Engineering, Tufts University, MA, USA. |
| Pseudocode | Yes | Algorithm 1 Iterative Refinement |
| Open Source Code | No | The paper states "Our model is implemented in PyTorch (Paszke et al., 2017)." but does not provide a specific link or explicit statement from the authors about releasing the source code for SIGMA. |
| Open Datasets | Yes | We use two datasets in this task, and follow the experiment setting from Xu et al. (2019a). In the first dataset, we use a Barabasi-Albert (BA) model to generate graphs of {500, 1000, 2000} nodes... In the second dataset, we start from the Protein-Protein Interaction (PPI) network of yeast (1,004 proteins and 4,920 interactions), and align its 5% noisy version provided in Saraph and Milenković (2014). For RDM pattern matching, "In the KEGG database". For image keypoints matching: "match keypoints on PASCAL VOC (Everingham et al., 2010) with Berkeley keypoint annotations (Bourdev and Malik, 2009)."... "SPair-71k (Min et al., 2019)." |
| Dataset Splits | Yes | We split the dataset into training, validation, testing at ratio 8:1:1. |
| Hardware Specification | Yes | Each model runs on a server with 32 cores and an NVIDIA A100 (40GB) GPU. |
| Software Dependencies | No | The paper states "Our model is implemented in PyTorch (Paszke et al., 2017)." and references other papers for baselines, but it does not specify explicit version numbers for the software libraries or frameworks used (e.g., PyTorch version, Python version, CUDA version). |
| Experiment Setup | Yes | For our model, we instantiate the GNN as a 5-layer Graph Isomorphism Network (GIN) (Xu et al., 2018). Each layer of GIN has a one-layer MLP with hidden dimension of 256 followed by a tanh( ) activation. The model is optimized by an Adam optimizer (Kingma and Ba, 2014) at a learning rate 10^-4 and trains for 100 epochs. For each dataset, the epoch that produces the best objective is used for testing. We use 10 samples of M. T is set to 4 (1 initial prediction followed by 4 iterations of refinement). |