Skip to content

Latest commit

 

History

History
129 lines (95 loc) · 4.01 KB

README.md

File metadata and controls

129 lines (95 loc) · 4.01 KB
Raw

Q-Lime banner

Q-LIME $\pi$: Quantum LIME for Text Explanations

This repository contains the implementation of Q-LIME $\pi$ (Quantum LIME), a quantum-inspired extension of LIME (Local Interpretable Model-agnostic Explanations), applied to text-based machine learning tasks. Q-LIME leverages quantum-inspired methods to explore the impact of flipping binary features on a model's output more efficiently.

Features

1. Classical and Quantum LIME Comparison

  • Classical LIME: Perturbs features to evaluate their impact on model predictions using a standard implementation.
  • Q-LIME $\pi$: Uses quantum-inspired binary encoding and flipping to compute feature importance.

2. Feature Visualization

  • Horizontal bar graph visualization for feature contributions.
  • Color-coded feature importance visualization directly on the text using HTML in Jupyter Notebooks.

3. Performance Evaluation

  • Benchmark comparisons between Classical LIME and Q-LIME $\pi$.
  • Metrics include runtime, accuracy, and overlap in top features.

This folder contains the exploratory work of Q-LIME $\pi$. For python package visit: https://github.com/nelabdiel/q_lime_package/

Installation

Prerequisites

  • Python 3.11+
  • Libraries:
    • numpy
    • pennylane
    • scikit-learn
    • matplotlib
    • lime
    • IPython

Install the required libraries using:

pip install numpy pennylane scikit-learn matplotlib lime ipython

Dataset

This project uses the IMDb sentiment analysis dataset. Download the dataset from IMDb Reviews and place it in the ./aclImdb/ directory.

Usage

Running the Main Script

The main.py script acts as the central entry point for the project, allowing users to execute key functionalities through a command-line interface (CLI). Choose from the following tasks:

  1. Run Q-LIME Example Visualization:

    • Visualizes feature contributions for a single text sample with a horizontal bar graph and highlights top contributing words directly on the text.
    python main.py example

  2. Run Classical vs Quantum LIME Comparison:

    • Compares runtime and feature overlap between Classical LIME and Q-LIME $\pi$.
    python main.py comparison

  3. Run Benchmark Tests:

    • Evaluates runtime, accuracy, and feature overlap across multiple configurations.
    python main.py benchmark

Script-Specific Functionality

highlighting.py

Contains the implementation of Q-LIME with:

  • Horizontal bar graph visualization of feature contributions.
  • Color-coded text highlighting for top features in Jupyter Notebook.

comparison.py

Compares Classical LIME and Q-LIME:

  • Measures runtime and top feature overlap.
  • Prints top-5 features for both methods.

benchmark.py

Benchmarks the performance of Classical LIME and Q-LIME:

  • Reports metrics for varying feature dimensions and dataset sizes.

Example Output

Q-LIME Visualization

Bar Graph for Feature Contributions (max_features=15)
Bar Graph Example

Text Highlighting Highlighted text with top contributing words color-coded (green for positive, red for negative).

Benchmark Results

Benchmark Test Output:

=== BENCHMARK RESULTS (Flip Only 1->0) ===
max_feats | Acc  | LIME_time | QLIME_time | Overlap
        5 | 0.450 | 1.126   | 0.005      | 2.50
       10 | 0.580 | 1.130   | 0.010      | 3.10
       15 | 0.600 | 1.079   | 0.075      | 2.90
       20 | 0.610 | 1.464   | 0.807      | 3.40

This benchmark demonstrates the trade-offs in runtime and overlap as the feature size increases.

Citation

If you use this repository in your work, please cite as:

@article{qlime2024,
  title={Q-LIME $\pi$: Quantum LIME for Text Explanations},
  author={Nelson Colon Vargas},
  year={2024},
  url={https://doi.org/10.48550/arXiv.2412.17197}
}

License

This project is licensed under the MIT License. See the LICENSE file for details.