NetGAN without GAN: From Random Walks to Low-Rank Approximations
Authors: Luca Rendsburg, Holger Heidrich, Ulrike Von Luxburg
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6. Experiments The purpose of this section is to (i) verify that CELL has performance comparable to Net GAN while being much faster, and (ii) demonstrate the importance of the cross-entropy loss and benefit of the logit-transformation by comparing with other low-rank approximation baselines. |
| Researcher Affiliation | Academia | 1Department of Computer Science, University of Tübingen, Germany 2Max Planck Institute for Intelligent Systems, Tübingen, Germany. |
| Pseudocode | Yes | Algorithm 1 Cross-Entropy Low-rank Logits (CELL) |
| Open Source Code | Yes | 1Code available at https://github.com/hheidrich/CELL |
| Open Datasets | Yes | Data sets and preprocessing. We experiment on a variety of graph data sets: the citation networks CORA-ML (Mc Callum et al., 2000) and CITESEER (Sen et al., 2008), the political blogs network POLBLOGS (Adamic & Glance, 2005), the retweet network RT-GOP, and the web graph WEB-EDU (Gleich et al., 2004). |
| Dataset Splits | Yes | For evaluating the link prediction performance during and after training, we split each graph into training-, validation-, and test-set by taking out 10% of the edges for validaton and another 5% for testing, while ensuring that the remaining graph stays connected. |
| Hardware Specification | No | Net GAN requires a GPU for training, while CELL runs on a CPU. |
| Software Dependencies | No | For solving optimization problem (10), we factorize W Wdown Wup with Wdown P RNˆH and Wup P RHˆN to satisfy the rank constraint, and optimize with Adam (Kingma & Ba, 2014). |
| Experiment Setup | Yes | To make the results comparable, we train CELL and Net GAN until the same stopping criterion of 52% edge overlap with the input graph is satisfied... using the logit space still has the advantage of requiring only a small rank (H 9 for CELL as compared to H 950 for LR-CE), which results in less trainable parameters and shorter training time. |