Skip to content

Latest commit

 

History

History
106 lines (73 loc) · 6.65 KB

README.md

File metadata and controls

106 lines (73 loc) · 6.65 KB

Interpretable Machine Learning for TabPFN

This repo contains all scripts for the standalone TabPFN interpretable ML package.

The project revolves around TabPFN with the corresponding paper.

This package is an extensive and easy-to-use toolbox of interpretability methods to be used in conjunction with TabPFN. It aims at leveraging TabPFN, which inherently possesses high predictive power, to explain its predictions.

To account for the peculiarities of TabPFN, many methods are implemented from scratch, as a significant speedup relative to existing open-source implementations could be achieved. For some methods, this package relies on packages by others, in particular:

  • shap by Scott Lundberg and collaborators.
  • dcurves by Daniel Sjoberg and collaborators.
  • MAPIE by Vianney Taquet, Grégoire Martinon and collaborators.

Moreover, within tabpfniml/tabpfn_interpret, it contains a modified copy of the TabPFN code. No modifications have been done that affect the inner workings of TabPFN or the predictions, however, in order to extract gradient information for our IML methods, the preprocessing steps (not differentiable in the original implementation) have been adapted and hooks have been added.

A demo notebook, containing examples of all methods available in this package, is located at demo/tabpfniml_demo_notebook.ipynb.

This package provides the following methods:

Feature Effect Methods

Method Name Filepath to Implementation Scope
Individual Conditional Expectation (ICE) curves tabpfniml/methods/ice_pd.py Local
Partial Dependence (PD) plots tabpfniml/methods/ice_pd.py Global
Accumulated Local Effects (ALE) plots tabpfniml/methods/ale.py Local
Kernel SHAP tabpfniml/methods/kernel_shap.py & tabpfniml/methods/kernel_shap_package.py Local & Global
Sensitivity Analysis tabpfniml/methods/sensitivity.py Local & Global

Feature Importance Methods

Method Name Filepath to Implementation Scope
Leave-One-Covariate-Out (LOCO) tabpfniml/methods/loco.py Global
Shapley Additive Global Importances (SAGE) tabpfniml/methods/kernel_shap.py Global

Data Valuation Methods

Method Name Filepath to Implementation Scope
Leave-One-Out (LOO) tabpfniml/methods/loco.py Global
Sensitivity Analysis tabpfniml/methods/sensitivity.py Global
Data Shapley tabpfniml/methods/data_shapley.py Global

Local Explanations

Method Name Filepath to Implementation Scope
Counterfactual Explanations (CE) tabpfniml/methods/counterfactuals.py Local

Clinical value assessment & Uncertainty quantification

Method Name Filepath to Implementation Scope
Decision Curve Analysis (DCA) tabpfniml/methods/dca.py Global
Conformal Prediction tabpfniml/methods/conformal_pred.py Local

For a detailed discussion of other available Transformer IML methods and their shortcomings w.r.t. TabPFN, refer to this markdown file.

Setup

To test this package, create a new conda environment. To do so, change the environment name package_test to a name of your choice and run the following:

conda create --name package_test python=3.8
conda activate package_test

Now, install the package tabpfn_iml from the local build file using pip into the conda environment:

pip install dist/tabpfniml-0.0.1.tar.gz

If it doesn't work (due to Python PATH conflicts), use:

/Users/<username>/miniconda3/envs/package_test/bin/pip install /Users/<username>/<path_to_repo>/tabpfn_iml/dist/tabpfniml-0.0.1.tar.gz

After successful installation, you can import methods from the package like this:

from tabpfniml import ICE_PD

Datasets

The IML methods can be applied to any dataset from the OpenML Benchmarking Suites using the OpenMLData class. Alternatively, the convenience class ArrayData can be used to construct a dataset instance from any array that can be transformed into a pandas DataFrame. Moreover, a new dataset instance that inherits from the dataset_iml class can be instantiated. Essentially, you only need to define the __init__() function. Please refer to tabpfniml/datasets/datasets.py for the structure and necessary variables. The datasets needs to be a .csv file and the first columns needs to be the target and the other colums the features. Categorical features should have already been converted to integer or float values, e.g., by label encoding.

from tabpfniml import dataset_iml

class myOwnData(dataset_iml):
  def __init__(self):
    ...

Experiments

All experiments conducted in the paper can be found in this repository, enabling the straightforward replication.

Experiment Paper Section Directory Comments
ICE & PD 4.1 experiments/ice_pd OpenML dataset ids: 11, 15, 31, 1049, 37, 40982, 1494
Kernel SHAP 4.2 experiments/kernel_shap
Context Optimization 4.3 experiments/context_optimization