wolfie00/surrogate-mtl

Implementation of "Exploring Multi-Task Learning for Explainability" in TensorFlow.

surrogate-mtl

Source code for the paper "Exploring Multi-Task Learning for Explainability" presented at ECAI's XI-ML 2023 workshop.

Dependencies

• TensorFlow 2/Keras
• TensorFlow Addons
• scikit-learn
• NumPy
• Pandas
• tqdm
• lime (https://github.com/marcotcr/lime/tree/master)
• If you want to create plots you will also need
  • pickle
  • matplotlib
  • paretoset (https://github.com/tommyod/paretoset)

Datasets

Experiments were conducted using some datasets from UCI such as Adult, AutoMPG and Red Wine Quality as well as California Housing and Titanic datasets.

Description

File	Content description
models.py	Implementation of STL baselines and MTL-based model classes in tf.keras.
trainer.py	Implementation of a Trainer class used to train the models.
evaluator.py	Implementation of an Evaluator class used to evaluate models using several metrics.
experiment.py	Class containing the logic for conducting the experiments.
main.py	Class that parses command-line arguments and calls Experiment from experiment.py.
data_files	Data files for the Adult dataset.

Usage

usage: main.py [-h] [-dataset {adult,wine,titanic,autompg}]
               [-stl_epochs STL_EPOCHS] [-mtl_epochs MTL_EPOCHS] [-regression]
               [-runs RUNS] [-es_patience ES_PATIENCE]
               [-pl_patience PL_PATIENCE] [-verbose] [-tune_arch]
               [-show_full_scores] [-save_plots]

optional arguments:
  -h, --help            show this help message and exit
  -dataset {adult,wine,titanic,autompg}
                        Name of dataset.
  -stl_epochs STL_EPOCHS
                        Number of STL training epochs.
  -mtl_epochs MTL_EPOCHS
                        Number of MTL training epochs.
  -regression           Whether the task is regression or (binary)
                        classification.
  -runs RUNS            Number of runs.
  -es_patience ES_PATIENCE
                        Early Stopping patience.
  -pl_patience PL_PATIENCE
                        Reduce learning rate on plateau patience.
  -verbose              Print training process.
  -tune_arch            Tune the MLP architecture.
  -show_full_scores     Prints a Pandas DataFrame with multiple scores in the
                        MTL setting.
  -save_plots           Whether to save plots of Accuracy/MSE-Fidelity.

For example:

python main.py -show_full_scores -runs=1 -dataset='adult'