Data Impact On XAI

Repository of the work entitled "Understanding the Influence of Data Characteristics on Explainable AI"

Structure

This repository has the following structure: ├── src/datasets ├── src/plots ├── src/results ├── src/dataset_descriptors.py ├── src/dataset_setup.py ├── src/dataset_utils.py ├── src/dnn_model.py ├── src/experiment.py ├── src/model_utils.py ├── src/results_analyzer.py ├── src/xai_methods.py

• src/datasets/ contains the datasets and its sources.
• src/plots/ contains all plots generated by the script src/results_analyzers.py
• src/results/ contains all logs from the experimental process
• src/dataset_descriptors.py contains the dataset descriptors
• src/dataset_setup defines the datasets to be used in the experiments
• src/dataset_utils.py contains the dataset loading and preprocessing functions
• src/dnn_model.py has the Deep Neural Networks code
• src/experiment.py has the main code of the experiment
• src/model_utils.py contains auxiliary methods to create and evaluate the models
• src/results_analyzer.py contains the script for generating the plots
• src/xai_methods.py contains the XAI methods code

Hyperparameter space explored for the ML models

| Model | Hyperparameters | |------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | LR | C: {0.1, 1, 10}
penalty: {l1, l2, None}
classweight: {None, balanced} | | DT | criterion: {gini, entropy, logloss}
maxdepth: {None, 10, 15, 20}
maxfeatures: {None, sqrt, log2}
classweight: {None, balanced} | | ExtraTree | criterion: {gini, entropy, logloss}
maxfeatures: {None, sqrt, log2}
maxdepth: {None, 10, 15, 20}
classweight: {None, balanced} | | SVM | kernel: {poly, rbf}
C: {0.1, 1, 10}
classweight: {None, balanced} | | GaussianNB | varsmoothing: {1e-12, 1e-10, 1e-8, 1e-6, 1e-4, 1e-2, 1} | | BernoulliNB | alpha: {0.1, 1, 10} | | RF | criterion: {gini, entropy, logloss}
maxfeatures: {None, sqrt, log2}
classweight: {None, balanced}
nestimators: {25, 50, 100, 200}
maxdepth: {None, 10, 15, 20} | | ExtraTrees | criterion: {gini, entropy, logloss}
maxfeatures: {None, sqrt, log2}
classweight: {None, balanced}
nestimators: {25, 50, 100, 200}
maxdepth: {None, 10, 15, 20} | | AdaBoost | nestimators: {25, 50, 100, 200}
learningrate: {0.1, 0.5, 1} | | GradientBoosting | criterion: {friedmanmse, squarederror}
nestimators: {50, 100, 200}
learningrate: {0.1, 0.5, 1}
maxdepth: {None, 10, 15, 20}
maxfeatures: {None, sqrt, log2} | | Bagging | nestimators: {10, 50, 100}
maxsamples: {1.0} | | MLP | hiddenlayersizes: {(50,), (100,)}
activation: {logistic, relu}
alpha: {0.0001, 0.001, 0.01} | | DNN | hiddenlayerssize: {16, 16, 16; 8, 16, 8; 32, 16, 8}
activationfunction: {sigmoid, tanh, relu, silu}
lossfunction: {categoricalcrossentropy}
nnoptimizer: {Adam, RMSprop, Nadam}
epochs: {100, 500}
batchsize: {10, 50} |

Summary of Best Hyperparameter Settings