Non-Euclidean Mixture Model for Social Network Embedding
Authors: Roshni Iyer, Yewen Wang, Wei Wang, Yizhou Sun
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on public datasets show NMM-GNN significantly outperforms state-of-the-art baselines on social network generation and classification tasks, demonstrating its ability to better explain how the social network is formed. |
| Researcher Affiliation | Academia | Roshni G. Iyer Computer Science Department University of California, Los Angeles Los Angeles, California, USA roshnigiyer@cs.ucla.edu Yewen Wang Computer Science Department University of California, Los Angeles Los Angeles, California, USA wyw10804@gmail.com Wei Wang Computer Science Department University of California, Los Angeles Los Angeles, California, USA weiwang@cs.ucla.edu Yizhou Sun Computer Science Department University of California, Los Angeles Los Angeles, California, USA yzsun@cs.ucla.edu |
| Pseudocode | No | The paper describes methods in text and uses tables (Table 3 and 4) for general frameworks, but does not provide explicit pseudocode or an algorithm block. |
| Open Source Code | Yes | Source code is in the Appendix. For details on future directions, the reader is also referred to the Limitations section in the Appendix. The source code and datasets for our work can be found at: https://github.com/roshnigiyer/nmm . |
| Open Datasets | Yes | Experiments on public datasets show NMM-GNN significantly outperforms state-of-the-art baselines on social network generation and classification tasks, demonstrating its ability to better explain how the social network is formed. For comprehensive evaluation, we assess our models on real-world datasets from well-known social media venues: Blog Catalog (BC) [27], Live Journal (LJ) [28], and Friendster (F) [29] which are friendship networks among bloggers. |
| Dataset Splits | Yes | For all experiments, for fairness of comparison to baselines, we utilize the experiment procedure of [6]. Specifically, 90% of links are randomly sampled as training data. We do not perform cross-validation, since it may cause overfitting to occur as our framework uses learnable parameters e.g., z S, z H, J, B, C, D, γ, β, λ, α, Wl, and ζ which is a function of z H equivalently interpreted as mean square error. Per dataset, we choose hyperparameter values for λA in reconstruction loss: {0, 1, 2, 4, 8, 16, 32, 64}, step sizes ηt: {0.005, 0.001, 0.01, 0.05, 0.1}, and experiments are performed on AWS cluster (8 Nvidia GPUs). Further, 10% of non-training edges are used for validation. |
| Hardware Specification | Yes | Per dataset, we choose hyperparameter values for λA in reconstruction loss: {0, 1, 2, 4, 8, 16, 32, 64}, step sizes ηt: {0.005, 0.001, 0.01, 0.05, 0.1}, and experiments are performed on AWS cluster (8 Nvidia GPUs). In compute, while we evaluate on a cluster of 8 GPUs, at least one GPU is necessary for running our experiments (which may been seen as compute limitation). |
| Software Dependencies | No | The paper does not provide specific software names with version numbers. |
| Experiment Setup | Yes | Per dataset, we choose hyperparameter values for λA in reconstruction loss: {0, 1, 2, 4, 8, 16, 32, 64}, step sizes ηt: {0.005, 0.001, 0.01, 0.05, 0.1}, and experiments are performed on AWS cluster (8 Nvidia GPUs). using embedding dimension 64. |