LIME — Explain Any Machine Learning Prediction

Introduction

LIME is a model-agnostic explanation technique introduced in the 2016 paper "Why Should I Trust You?" by Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. It explains individual predictions by perturbing the input and fitting a simple, interpretable model around the neighborhood of each data point. LIME works with any black-box classifier or regressor.

What LIME Does

• Explains individual predictions for tabular, text, and image data
• Generates human-readable feature importance for any model
• Works with scikit-learn, XGBoost, TensorFlow, PyTorch, and any predict function
• Highlights which input features drove a specific prediction
• Supports submodular pick (SP-LIME) for selecting representative explanations

Architecture Overview

LIME perturbs the input around a data point, obtains predictions from the black-box model for each perturbation, and fits a weighted linear model locally. The weights decrease with distance from the original point. For text, it removes words; for images, it masks superpixels; for tabular data, it samples from the training distribution. The resulting sparse linear model coefficients become feature importances.

Self-Hosting & Configuration

• Install via pip: pip install lime
• Requires Python 3.6+ with NumPy and scikit-learn
• No external services or APIs needed
• Works in Jupyter notebooks with built-in HTML visualizations
• Configurable: number of features, perturbation count, and kernel width

Key Features

• Truly model-agnostic: works with any classifier that exposes a predict function
• Built-in support for tabular, text, and image explanation modes
• Generates interactive HTML explanations for notebooks
• SP-LIME picks a diverse set of representative explanations for a dataset
• Lightweight with minimal dependencies

Comparison with Similar Tools

• SHAP — provides global and local explanations with Shapley values; LIME is faster for single predictions
• Captum — PyTorch-specific attribution methods; LIME is framework-agnostic
• InterpretML — Microsoft's unified interpretability; LIME focuses on local explanations
• Anchor — by the same authors, provides rule-based explanations with coverage guarantees

FAQ

Q: How is LIME different from SHAP? A: LIME fits a local linear model around each prediction. SHAP computes Shapley values, offering stronger theoretical guarantees but at higher computational cost for some model types.

Q: Can LIME explain deep learning models? A: Yes. As long as the model exposes a prediction function, LIME can explain it regardless of architecture.

Q: Is LIME suitable for production use? A: LIME is commonly used in production for model auditing and compliance. Explanation latency depends on the number of perturbations configured.

Q: Does LIME support regression? A: Yes. Use LimeTabularExplainer with mode set to regression.

Sources

• https://github.com/marcotcr/lime
• https://arxiv.org/abs/1602.04938