Deep Relational Topic Modeling via Graph Poisson Gamma Belief Network
Authors: Chaojie Wang, Hao Zhang, Bo Chen, Dongsheng Wang, Zhengjue Wang, Mingyuan Zhou
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our models extract high-quality hierarchical latent document representations, leading to improved performance over baselines on various graph analytic tasks. |
| Researcher Affiliation | Academia | Chaojie Wang , Hao Zhang , Bo Chen , Dongsheng Wang, Zhengjue Wang National Laboratory of Radar Signal Processing Xidian University, Xi an, Shaanxi 710071, China xd_silly@163.com, zhanghao_xidian@163.com, bchen@mail.xidian.edu.cn wds_dana@163.com, zhengjuewang@163.com Mingyuan Zhou Mc Combs School of Business The University of Texas at Austin Austin, TX 78712, USA mingyuan.zhou@mccombs.utexas.edu |
| Pseudocode | Yes | The detailed training algorithm is provided in Appendix C, and the released code3 is implemented with Tensor Flow [47], combined with py CUDA [48] for parallel Gibbs sampling. |
| Open Source Code | Yes | The detailed training algorithm is provided in Appendix C, and the released code3 is implemented with Tensor Flow [47], combined with py CUDA [48] for parallel Gibbs sampling. |
| Open Datasets | Yes | We consider six widely used benchmarks, including Coil [5], TREC [43], and R8 [49] for node clustering, and Cora, Citeseer and Pubmed [50] for link prediction and node classification. |
| Dataset Splits | Yes | Following VGAE [23], we train the model on an incomplete version of the network data, with 5% and 10% of the citation links used for validation and test, respectively. |
| Hardware Specification | No | The paper mentions acceleration with 'GPU' but does not provide specific hardware details such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The released code is implemented with Tensor Flow [47], combined with py CUDA [48] for parallel Gibbs sampling. |
| Experiment Setup | Yes | We perform three WGAEs/WGCAEs with different stochastic layers, i.e., T {1, 2, 3}, and set the network structure as K1 = K2 = K3 = C, where C is set as the total number of classes for node clustering/classification, and 16 for link prediction following VGAE [23] to make a fair comparision. In Fig. 4, we can see that the best performance of node clustering and link prediction are achieved around β = 0.1 and β = 100, respectively. |