Causal Discovery and Knowledge Injection for Contestable Neural Networks

This repository provides the code for the paper on Contestable Neural Networks.

There are three main python scripts:

• net_inject.py defines the InjectedNet class: this is used to apply the causal masking to the weights of the input layer of the neural network, enforcing the use of only the direct causal relationships provided in the input DAG/graph.

• main_synth.py reproduces the experiments on synthetic data. The script iterates over the three main dimensions varying for the generated DAGs: number of nodes in DAG (|V| in {10, 20, 50}), number of edges (|E| = |V| x e, where e in {1,2,5}), and data size (N=|V| x s, where s in {50,100,200,300,500}).

• main_realdata.py reproduces the experiments on real data. The experiments on real data are on FICO HELOC, Adult, Boston and California Housing datasets. The script iterates over two dimensions: the threshold tau (to retrieve computed DAGs) and the sample size N. All the data is downlowded within the utils.py script apart from HELOC (this is instead provided in the fico folder within data directory, both pre-processed and original).

Example usage

To fit an injected network, from python at the root folder, run:

# Imports
from sklearn.model_selection import train_test_split
import numpy as np
import pandas as pd
from net_inject import InjectedNet
from utils import load_or_gen_data, load_adj_mat

# Load toy data
G, df = load_or_gen_data(name='toy', N=5000, csv='data/synth_nonlinear.csv')
loaded_adj = load_adj_mat(df, DAG_inject='toy', version='test')
X_DAG = df.to_numpy()
y_DAG = np.expand_dims(X_DAG[:,0], -1)
train_X, test_X, train_y, test_y = train_test_split(X_DAG, y_DAG, test_size=0.2, random_state=0)

# Fit toy model
net = InjectedNet(num_inputs = X_DAG.shape[1])
net.fit(X=train_X, y=train_y, num_nodes=X_DAG.shape[1], X_val=test_X, y_val=test_y, inject=True, verbose=True)

To rerun the synthetic experiments, from a terminal at the root folder, run:

python main_synth.py --version="r_ex_1b_k20_l3_h2_n0" --branchf=1 --known_p=0.2 --noise_p=0.0 --hidden_l=3 --hidden_n_p=2

main_realdata.py has the same usage. Refer to the scripts for descriptions of the meaning and usage of the parameters. Beware, an end-to-end run of this script takes several hours if run on single cpu/gpu.

Environment

The code was tested with Python 3.6.9 and 3.9.9. requirements.txt provides the necessary packages. Run pip install -r requirements.txt from a terminal at the root folder to install all packages in your virtual environment.

Reference

If you are using this code, please cite our paper

@inproceedings{Russo2023contestable,
  author       = {Fabrizio Russo and
                  Francesca Toni},
  title        = {Causal Discovery and Knowledge Injection for Contestable Neural Networks},
  booktitle    = {{ECAI} 2023 - 26th European Conference on Artificial Intelligence,
                  September 30 - October 4, 2023, Krak{\'{o}}w, Poland - Including
                  12th Conference on Prestigious Applications of Intelligent Systems
                  {(PAIS} 2023)},
  series       = {Frontiers in Artificial Intelligence and Applications},
  volume       = {372},
  pages        = {2025--2032},
  publisher    = {{IOS} Press},
  year         = {2023},
  url          = {https://doi.org/10.3233/FAIA230495},
  doi          = {10.3233/FAIA230495},
}