Exploring the Cloud of Feature Interaction Scores in a Rashomon Set

This repository includes the implementation of the paper (https://arxiv.org/abs/2305.10181). The paper emphasizes the importance of investigating feature interactions not just in a single predictive model, but across a range of well-performing models, illustrated below.

Summary

The paper introduces a novel approach to understanding feature interactions in machine learning models. The authors argue that the study of feature interactions in a single model can miss complex relationships between features. Thus, they recommend exploring these interactions across a set of similarly accurate models, known as a Rashomon set.

The main contributions of this paper are:

1. Introduction of the Feature Interaction Score (FIS) as a means to quantify the strength of feature interactions within a predictive model.
2. Proposal of the FIS Cloud (FISC), a collection of FISs across a Rashomon set, to explore how feature interactions contribute to predictions and how these interactions can vary across models.
3. Presentation of a non-linear example in a multilayer perceptron (MLP) to characterize the Rashomon set and FISC. This includes a search algorithm to explore FIS and to characterize FISC in a Rashomon set along with two novel visualization tools, namely Halo and Swarm plots.

---

Project structure

The project is constructed following the below structure:

project
│   README.md
│   requirement.txt    
│   LICENSE
└───data
│   │   data_file.csv
└───grs
│   │───explainers
│   │   │    __init__.py
│   │   │    _explainer.py
│   │───plots
│   │   │    __init__.py
│   │   │    _swarm_plot.py
│   │   │    _halo_plot.py
│   │───utils
│   │   │    __init__.py
│   │   │    _general_utils.py
│   │   __init__.py
│   │   config.py
└───demo
│   │   demo.ipynb
└───experiments
│   │   ...
└───logs
│   │   ...
└───results
│   │   ...
───────────

Requirements

FISC is a designed based on Python language and specific libraries. Four most commonly-used are listed as follows:

• Python
• Jupyter Notebook
• Numpy
• Matplotlib

FISC is a model-agnostic framework and experiments might have different requirements, depending on the data and model types. To implement the experiments in the paper, dependencies in requirements.txt are required.

python -r .\requirements

---

Installment

pip install -i https://test.pypi.org/simple/ generalized-rashomon-set

Usage

from generalized_rashomon_set import utils
from generalized_rashomon_set import plots
from generalized_rashomon_set import explainers
from generalized_rashomon_set import config


# train a MLP model
model =  MLPRegressor(hidden_layer_sizes, max_iter=200).fit(X, y)

# explain the model by exploring the Rashomon set
explainer = explainers.fis_explainer(model, X_test, y_test, epsilon_rate=0.05,
                                     loss_fn='mean_squared_error', n_order=2, delta=0.5, torch_input=False)
explainer.ref_explain()
explainer.rset_explain()

# visualize the predictions
plots.swarm_plot_MR()
plots.swarm_plot_FIS()
plots.halo_plot()

Key parameters

• model: the trained reference/optimal model, compatible with sklearn and pytorch models
• input: the input data, a numpy array or a torch tensor
• output: the output data, a numpy array or a torch tensor
• epsilon_rate: a positive value used to set the bound for the Rashomon set, default is 0.1
• loss_fn: the loss function, default is 'mean_squared_error' for regression and 'log_loss' for classification
• n_order: the number to define the highest order of attribution, default is 2, which means the feature interaction
• wrapper_for_torch: a boolean value to indicate whether the model is a pytorch model, set to False when using sklearn models, default is False
• delta: a positive value to set the step size of searching the Rashomon set, default is 0.1; Note: the current version only supports delta=0.1 for interaction plots
• binary: a boolean value to indicate whether the loss is calculated by scores or binary classification results, default is False and only used for classification tasks

---

Demo

A toy example shows how halo plot illustrates the effect of feature interaction in toy-example.ipynb, results are shown below.

Experiment

A more detailed example of recidivism prediction for real-world applications can be found in experiments recidivism prediction.ipynb.

The results might be slightly different from the paper due to the refactor of the code, but the main conclusions remain the same. The package is still under refactoring to facilitate easy visualization, and we are actively working on a more comprehensive documentation.

Please feel free to drop an email to Sichao Li in case you want to discuss.

Citation

@article{li2023exploring,
  title={Exploring the cloud of feature interaction scores in a Rashomon set},
  author={Li, Sichao and Wang, Rong and Deng, Quanling and Barnard, Amanda},
  journal={arXiv preprint arXiv:2305.10181},
  year={2023}
}