DeepDRK: Deep Dependency Regularized Knockoff for Feature Selection
Authors: Hongyu Shen, Yici Yan, Zhizhen Jane Zhao
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experiments conducted on synthetic, semi-synthetic, and real-world datasets demonstrate that our pipeline outperforms existing methods across various scenarios. |
| Researcher Affiliation | Academia | Department of Electrical and Computer Engineering1, University of Illinois at Urbana Champaign. Department of Statistics2, University of Illinois at Urbana Champaign. |
| Pseudocode | Yes | In Algorithm 1, we provide pseudo code for training the Knockoff Transformer and the swappers (i.e., the first stage shown in Figure 1). |
| Open Source Code | Yes | Deep DRK is implemented in Py Torch [45] and is accessible at: https://github.com/nowonder2000/Deep DRK. |
| Open Datasets | Yes | The first dataset contains single-cell RNA sequencing (sc RNA-seq) data from 10ˆ Genomics 7. ... The second publicly available dataset9 is from a real case study entitled Longitudinal Metabolomics of the Human Microbiome in Inflammatory Bowel Disease (IBD) [35]. |
| Dataset Splits | Yes | To fit models, we first split datasets of X into training and validation sets with an 8:2 ratio. The training sets are used for model optimization, and the validation sets are used for early stopping based on the validation loss, with a patience period of 6. |
| Hardware Specification | Yes | Experiments are conducted on a single NVIDIA V100 16GB GPU. |
| Software Dependencies | No | The paper states that 'Deep DRK is implemented in Py Torch [45]' but does not provide a specific version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | We follow the model configuration in Table 3 to optimize Deep DRK. The architecture of the swappers Sω is based on [55]. Both the swappers and Xθ are trained using the Adam W optimizer [36]. During training, we alternately optimize Xθ and the swappers Sω, updating weights θ three times for each update of weights ω. This training scheme is similar to GAN training [21], though without discriminators. We apply early stopping to prevent overfitting. A pseudocode of the optimization is provided in Appendix B.2. In experiments, we set αn 0.5 universally as the dependency regularization coefficient due to its consistent performance 16. A discussion on the effect of αn is provided in Appendix G. ... Parameter Value Sω Learning Rate 1 ˆ 10 3 Xθ Learning Rate 1 ˆ 10 5 Dropout Rate 0.1 # of Epochs 200 Batch Size 64 λ1 30.0 λ2 1.0 λ3 20.0 Early Stop Tolerance 6 αn 0.5 |