Topic Modeling on Document Networks with Adjacent-Encoder
Authors: Ce Zhang, Hady W. Lauw6737-6745
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our models on real-world document networks quantitatively and qualitatively, outperforming comparable baselines comprehensively. |
| Researcher Affiliation | Academia | Ce Zhang, Hady W. Lauw School of Information Systems Singapore Management University {cezhang.2018, hadywlauw}@smu.edu.sg |
| Pseudocode | No | The paper describes the model architecture with mathematical equations and accompanying text, but it does not include a formally labeled "Pseudocode" or "Algorithm" block. |
| Open Source Code | Yes | All data and code are available for reproducibility2. 2https://github.com/cezhang01/Adjacent-Encoder |
| Open Datasets | Yes | Cora1 is a public collection of papers and their citations (Mc Callum et al. 2000). Each document is an abstract. Two documents are linked by an undirected edge if one cites the other. Following (Zhu et al. 2007), we extract four independent datasets: Data Structure (DS), Hardware and Architecture (HA), Machine Learning (ML), and Programming Language (PL). 1http://people.cs.umass.edu/mccallum/data/cora-classify.tar.gz |
| Dataset Splits | Yes | In the inductive setting, the objective is to generalize beyond the training corpus to unseen data, which we simulate by keeping a random subset of 80% documents for training (out of which we further randomly split 10% documents for validation) and the remaining 20% for testing. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments, such as CPU models, GPU models, or memory specifications. |
| Software Dependencies | No | The paper mentions types of activation functions (sigmoid, tanh) and parameters for some models (e.g., DAE with Gaussian noise of 0.25 std.dev., Dirichlet hyperparameter for RTM and PLANE, number of nonzero hidden neurons for KSAE and KATE). However, it does not list specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, TensorFlow 2.x). |
| Experiment Setup | Yes | Following (Chen and Zaki 2017; Bai et al. 2018), the activation functions for AE, DAE, CAE, KSAE, and NRTM are sigmoid, while those for VAE and KATE are tanh (hidden) and sigmoid (output) respectively. We use validation set to choose the best hyperparameters. DAE with Gaussian noise of 0.25 std.dev. outperforms other kinds of noise. We choose 2 and 0.01 as Dirichlet hyperparameter for RTM and PLANE. For KSAE and KATE, we set the number of nonzero hidden neurons, k, to 4, 8, 16, 32, and 52 when the number of topics is 16, 32, 64, 128, and 256, respectively. |