NetGAN: Generating Graphs via Random Walks
Authors: Aleksandar Bojchevski, Oleksandr Shchur, Daniel Zügner, Stephan Günnemann
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section we evaluate the quality of the graphs generated by Net GAN by computing various graph statistics. We quantify the generalization power of the proposed model by evaluating its link prediction performance. |
| Researcher Affiliation | Academia | 1Technical University of Munich, Germany. Correspondence to: Daniel Z ugner <zuegnerd@in.tum.de>. |
| Pseudocode | Yes | The generative process of G is summarized in the box below. m0 = gθ (z) v1 Cat(σ(p1)), (p1, m1) = fθ(m0, 0) v2 Cat(σ(p2)), (p2, m2) = fθ(m1, v1) ... ... v T Cat(σ(p T )), (p T , m T ) = fθ(m T 1, v T 1) |
| Open Source Code | Yes | 1Code available at: https://www.kdd.in.tum.de/netgan |
| Open Datasets | Yes | Datasets. For the experiments we use five well-known citation datasets and the Political Blogs dataset. ... CORA-ML (Mc Callum et al., 2000) CITESEER (Sen et al., 2008) PUBMED (Sen et al., 2008) DBLP (Pan et al., 2016) POL. BLOGS (Adamic & Glance, 2005) |
| Dataset Splits | Yes | During training, we keep a sliding window of the random walks generated in the last 1,000 iterations and use them to construct a matrix of transition counts. This matrix is then used to evaluate the link prediction performance on a validation set (i.e. ROC and AP scores, for more details see Sec. 4.2)." and "We hold out 10% of edges from the graph for validation and 5% as the test set, along with the same amount of randomly selected non-edges, while ensuring that the training network remains connected. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions software components like LSTM, WGAN, and Adam (Kingma & Ba, 2014), but does not provide specific version numbers for any software dependencies used in their implementation or experiments. |
| Experiment Setup | Yes | The choice of τ allows to trade-off between better flow of gradients (large τ, more uniform v t ) and more exact calculations (small τ, v t vt). ... We train our model based on the Wasserstein GAN (WGAN) framework (Arjovsky et al., 2017), as it prevents mode collapse and leads to more stable training overall. To enforce the Lipschitz constraint of the discriminator, we use the gradient penalty as in Gulrajani et al. (2017). The model parameters {θ, θ } are trained using stochastic gradient descent with Adam (Kingma & Ba, 2014). Weights are regularized with an L2 penalty. ... One important exception is the the random walk length T. To choose the optimal value, we evaluate the change in link prediction performance as we vary T on CORAML. ... The performance gain for T = 20 over T = 16 is marginal and does not outweigh the additional computational cost, thus we set T = 16 for all experiments. |