Shapley Explanation Networks
Authors: Rui Wang, Xiaoqian Wang, David I. Inouye
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate on synthetic and real-world datasets that our SHAPNETs enable layer-wise Shapley explanations, novel Shapley regularizations during training, and fast computation while maintaining reasonable performance. Code is available at https: //github.com/inouye-lab/Shapley Explanation Networks. 1 INTRODUCTION Explaining the predictions of machine learning models has become increasingly important for many crucial applications such as healthcare, recidivism prediction, or loan assessment. ... 3 EXPERIMENTS & VISUALIZATIONS We will (1) validate that our SHAPNET models can be quite expressive despite the intrinsic explanation design, (2) demonstrate that our intrinsic SHAPNET explanations perform comparably or better than post-hoc explanations, and (3) highlight some novel capabilities. |
| Researcher Affiliation | Academia | Rui Wang Xiaoqian Wang David I. Inouye School of Electrical and Computer Engineering Purdue University West Lafayette, IN 47906 {ruiw, joywang, dinouye}@purdue.edu |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Code is available at https: //github.com/inouye-lab/Shapley Explanation Networks. |
| Open Datasets | Yes | First, we create a synthetic regression dataset... Second, we choose five real-world datasets from Dua & Graff (2017): Yeast (d = 8) and Breast Cancer Wisconsin (Diagnostic) (d = 30), MNIST (Le Cun & Cortes, 2010), Fashion MNIST (Xiao et al., 2017), and Cifar-10 (Krizhevsky, 2009) datasets to validate SHAPNET on higher-dimensional datasets. |
| Dataset Splits | Yes | For each round, we perform 5-fold cross validation on the training set, and arrive at 5 different models, giving us 50 different models to explain for each datasets. ... For the preprocessing procedures, we first randomly split the data into training and testing sets, with 75% and 25% of the data respectively. |
| Hardware Specification | Yes | The model of the CPU used for testing is Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz. |
| Software Dependencies | No | For all the experiments described below, the optimizers are Adam (Kingma & Ba, 2014) with default parameters in Py Torch (Paszke et al., 2019). The paper cites the PyTorch paper but does not provide a specific numerical version number for PyTorch or other software components. |
| Experiment Setup | Yes | As discussed before, all the models are trained with Adam in Py Torch with the default parameters. The batch size is set to 64 for all the vision tasks. No warm-up or any other scheduling techniques are applied. The Fashion MNIST & MNIST models in Table 1 are trained for 10 epochs while the MNIST models with which we investigate the new capabilities (layer-wise explanations & Shapley explanation regularizations) in Section 3 are trained with just 5 epochs. The Cifar-10 models are trained for 60 epochs. |