IML_for_Imbalanced_Learning

This repository is for the paper, "Interpretable ML for Imbalanced Data." It contains the code and links to obtain pre-trained models, as well as the steps, to reproduce several of the visualizations listed in the paper. Please note that the code provided below is for the CIFAR-10 dataset.

Steps

1. Extract FE from a trained model.

• Start with extract_FE.py
  • link to pre-trained model used by Extract_FE: https://drive.google.com/file/d/18yWsVlUNVrSMq4qTlY2d9MPHThZaSWUK/view?usp=sharing

2. Display class accuracy.

• class_accuracy.py

3. Visualize class archetypes (safe, border, rare, outliers) for a specific class

• First, generate nearest neighbors with arch_NNB.py
• Alternatively, use the CE_cif_trn_NNB.csv file in the data folder.
• Run k_medoid_viz.py

3. Visualize nearest adversary neighbors bar chart

• Run NNB_FP_bar.py

4. Display feature embedding (FE) top-10 indices and FE densities

• Run FE_idx_density.py

5. Visualize color bands for a specific class of interest

• Run saliency_texture.py