rehmanzafar/dlime_experiments

In this work, we propose a deterministic version of Local Interpretable Model Agnostic Explanations (LIME) and the experimental results on three different medical datasets shows the superiority for Deterministic Local Interpretable Model-Agnostic Explanations (DLIME).

Deterministic Local Interpretable Model-Agnostic Explanations for Stable Explainability

Experiments

Setup Environment

The following python environment and packages are used to conduct the experiments:

• python==3.6
• Boruta==0.1.5
• numpy==1.16.1
• pandas==0.24.2
• scikit-learn==0.20.2
• scipy==1.2.1

These packages can be installed by executing the following command: pip3.6 install -r requirements.txt

Datasets

To conduct the experiments we have used the following three healthcare datasets from UCI repository:

• Breast Cancer
• Hepatitis
• Indian Liver Patient
• Cardiotocography
• Thyroid
• Handwritten Digits

Breast cancer dataset comes along with scikit-learn package, therefore, there is no need to download this dataset. The rest of the datasets are already downloaded and available in "data" folder.

Algorithms

The following classifiers and algorithms are used in this study:

• Random Forest
• Neural Networks
• Linear Regression
• Logistic Regression
• K-Nearest Neighbours
• K-Means Clustering
• Agglomerative Hierarchical Clustering

Execute Code

Run the following files to reproduce the results. The results of LIME are not deterministic and it may produce different results.

Experiments on Breast Cancer Dataset:

• python3.6 experiments_bc_nn.py
• python3.6 experiments_bc_rf.py

Experiments on Indian Liver Patient Dataset:

• python3.6 experiments_ildp_nn.py
• python3.6 experiments_ildp_rf.py

Experiments on Hepatitis Dataset:

• python3.6 experiments_hp_nn.py
• python3.6 experiments_hp_rf.py

For the quality of the explanations:

• python3.6 experiments_bc_lgr_fidelity_v2p0-mc-v2.py
• python3.6 evaluate_quality_v0.py

Results

The results will be saved inside "results" directory in pdf and csv format. The quality of the explanation is shown in the image below:

Citation

Please consider citing our work if you use this code for your research.

Initial Results

@InProceedings{zafar2019dlime,
  author    = {Muhammad Rehman Zafar and Naimul Mefraz Khan},
  title     = {DLIME: A Deterministic Local Interpretable Model-Agnostic Explanations Approach for Computer-Aided Diagnosis Systems},
  booktitle = {In proceeding of ACM SIGKDD Workshop on Explainable AI/ML (XAI) for Accountability, Fairness, and Transparency},
  year      = {2019},
  publisher = {ACM},
  address   = {Anchorage, Alaska}
}

Extended Version

@article{zafar2021deterministic,
  title={Deterministic Local Interpretable Model-Agnostic Explanations for Stable Explainability},
  author={Zafar, Muhammad Rehman and Khan, Naimul},
  journal={Machine Learning and Knowledge Extraction},
  volume={3},
  number={3},
  pages={525--541},
  year={2021},
  publisher={Multidisciplinary Digital Publishing Institute}
}

dlime_experiments's Issues

Support for python 3.6

Dear Sir,

Currently, there is no support for Python 3.6. So how can I use this code? I am interested in using code for project work. Urgently needed.

error when reproducing the results

Thank you for the great paper.

When I was running the code in experiments_bc_rf.py, I got an error in line 80

This method as_pyplot_to_figure belongs to Explanation object but not to explainer_tabular. LimeTabularExplainer. It seems to me that these two objects are independent from each other. Have I missed anything? Could you please advise how to reproduce the result correctly?

Here is the Colab I used.

Ask questions

Hi, thank you very much for sharing the code.
But there is a problem I want to ask, that is, after improvement, the contribution of each feature of the advanced interpretation result is the same. I don’t know whether such a result still has interpretive significance.

problem feedback