PieClam: A Universal Graph Autoencoder Based on Overlapping Inclusive and Exclusive Communities
Authors: Daniel Zilberg, Ron Levie
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 4 we support our construction with experiments, where our models achieve competitive performance with respect to state of the art. First, we use Pie Clam to perform graph anomaly detection. Here, Pie Clam learns a probabilistic model given graph, and this model can be used for inspecting the probabilities of different nodes in this graph: nodes with low probabilities are deemed to be anomalies. Then, we use Pie Clam to predict edges in link prediction benchmarks. We conducted additional link prediction experiments on the OGB-DDI dataset. |
| Researcher Affiliation | Academia | 1Faculty of Mathematics, Technion Israel Institute of Technology, Haifa, Israel. Correspondence to: Daniel Zilberg <EMAIL>, Ron Levie <EMAIL>. |
| Pseudocode | No | The paper describes methods and algorithms using mathematical notation and textual explanations, but it does not include any explicitly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | Our code can be found in: https://github.com/danizil/PieClam |
| Open Datasets | Yes | In Table 1 we compare the performance of Clam methods to DOMINANT (Ding et al., 2019a), Anomaly DAE (Fan et al., 2020), OCGNN (Wang et al., 2021), AEGIS (Ding et al., 2021), GAAN (Chen et al., 2020b) and TAM (Qiao & Pang, 2024) on the datasets Reddit, Elliptic (Elliptic) (Weber et al., 2019), and Photo (Shchur et al., 2018a). ... We run our experiments on the datasets Squirrel and Texas (Pei et al., 2020), Photo (Shchur et al., 2018b) and Facebook100 s Johns Hopkins (Lim et al., 2021b), and compare Pie Clam to the baselines AA (Adamic & Adar, 2003), VGAE (Kipf & Welling, 2016), GAT (Velickovic et al., 2017), LINKX (Lim et al., 2021a) and Disenlink (Zhou et al., 2022). ... We conducted additional link prediction experiments on the OGB-DDI dataset. |
| Dataset Splits | Yes | For each dataset, we generate 10 random splits into a test set containing 10% of the edges, along with 5 randomly sampled non-edges for every omitted edge, and the rest is the training set. For each split and each configuration of the hyperparameters, we generate three random validation sets consisting of 5% of the edges from the training set and 5 non-edges per omitted edge also from the training set, leaving the rest of the original training set for training. |
| Hardware Specification | Yes | All of the experiments were run on Nvidia Ge Force RTX 4090 and Nvidia L40 GPUs. |
| Software Dependencies | No | The paper does not explicitly state the specific version numbers for any software dependencies used in the implementation or experiments. It mentions various models and methods but no software versions (e.g., Python, PyTorch, TensorFlow, etc.). |
| Experiment Setup | Yes | We use an affiliation space embedding dimension of 30 for F for Pie Clam and Ie Clam, and 24 for Big Clam. Every Clam method starts with a random embedding F. ... We train Pie Clam with the scheme F-500 p-1300 F-500 p-1300 F-500 p-1300 with learning rate of 2e 6 on F and 1e 6 on p. In every alternation between two (p t) and (F t) schemes, both learning rates are decreased by a factor of 2. For the models which only optimize F (Ie Clam and Big Clam) we use the following configurations. We train Ie Clam with 2500 iterations with learning rate 1e-6. We train Big Clam with 2200 iterations with learning rate 1e-6. ... The following hyperparameters are constant across all datasets: feature step size (5e-6), l1 regularization (1), Noise amplitude (0.1) the number of alternations (7), scheduler step size (3) and the scheduler step size (0.5). |