Enhancing Counterfactual Classification Performance via Self-Training
Authors: Ruijiang Gao, Max Biggs, Wei Sun, Ligong Han6665-6673
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of the proposed algorithms on both synthetic and real datasets. |
| Researcher Affiliation | Collaboration | Ruijiang Gao*1, Max Biggs2, Wei Sun3, Ligong Han4 1 University of Texas at Austin 2 University of Virginia 3 IBM Research 4 Rutgers University |
| Pseudocode | Yes | Algorithm 1: Counterfactual Self-Training |
| Open Source Code | Yes | Code is available at https://github.com/ruijiang81/CST |
| Open Datasets | Yes | Multi-Label Datasets We use two multi-label datasets from LIBSVM repository (Elisseeff and Weston 2002; Boutell et al. 2004; Chang and Lin 2011), which are used for semantic scene and text classification. |
| Dataset Splits | No | The hyperparameter of CVAT loss is chosen using a grid search of [0.01, 0.1, 1, 10] on a validation set. The paper mentions using a 'validation set' for hyperparameter tuning but does not specify the size, percentage, or how this split is created for reproducibility. |
| Hardware Specification | Yes | All experiments are conducted using one NVidia GTX 1080-Ti GPU with five repetitions and Adam optimizer (Kingma and Ba 2014). |
| Software Dependencies | No | The paper mentions 'Adam optimizer' and 'cross entropy loss' but does not specify specific software or library versions (e.g., PyTorch, TensorFlow) needed for reproducibility. |
| Experiment Setup | Yes | For the implementation, we use a three layer neural network with 128 nodes as our model with dropout of p = 0.2 and Leaky Re LU activation unless specified otherwise, and cross entropy loss as the loss function. ... The stepsize in finite-difference approximation, number of power iteration steps and perturbation size are set to 10, 3 and 1 respectively. The number of iteration of CST is set to 2. The hyperparameter of CVAT loss is chosen using a grid search of [0.01, 0.1, 1, 10] on a validation set. |