Counterfactual WOA and GA

• Interpretation of Black-Box Model with Counterfactual Instance generated by Metaheuristics
  • used Whale Optimization Algorithm and Genetic Algorithm

Concept

• Counterfactual
• Generate the most similar input data (features) with the most different predictions
  • the more changes (of input features), the more important
• Objective Function
  • minimize difference between a sample (which we want to interprete) features and a generated one
  • maximize difference between a sample (which we want to interprete) targets and a generated one
  • I used RMSE but you can use Cosine Similarity considering your input data distribution

Data

• N_features : 10
  • Not Informative Features : 4, 6, 9 (index)
• N_targets : 3
• This data was generated by using sklearn's make_regression function

Model

• simple DNN
  • 5 fc layers

Requirements

• geneticalgorithm==1.0.2
• matplotlib==3.5.1
• numpy==1.20.0
• pandas==1.3.4
• scikit-learn==1.0.1
• scipy==1.7.3
• torch==1.10.0

pip install -r requirements.txt