# "Why Should I Trust You?": Explaining the Predictions of Any Classifier

**Authors:** Marco Tulio Ribeiro, Sameer Singh, Carlos Guestrin  
**arXiv:** [1602.04938](https://arxiv.org/abs/1602.04938)  
**Date:** 2016-02-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1602.04938v3/

## Key points

- Machine learning models are often considered black boxes, leading to a lack of trust in their predictions.
- Understanding the rationale behind predictions is crucial for establishing trust and making informed decisions.
- LIME is introduced as a novel explanation technique to elucidate predictions of any classifier in an interpretable manner by constructing local interpretable models around each prediction.
- The authors propose a method for explaining models by showcasing representative individual predictions along with explanations in a non-redundant way, treating it as a submodular optimization problem.
- Experiments demonstrate the flexibility and applicability of these methods across different types of models for text and image classification.
- The utility of explanations is highlighted through scenarios like deciding whether to trust a prediction, selecting between models, improving an unreliable classifier, and identifying reasons why a classifier should not be trusted.

## Summary

, , , , 

In the paper titled "Why Should I Trust You? ": Explaining the Predictions of Any Classifier, authors Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin address the issue of machine learning models being predominantly black boxes despite their widespread adoption. They highlight the importance of understanding the rationale behind predictions to establish trust, especially when making decisions based on these predictions or considering deploying a new model. The authors emphasize that gaining insights into the model can help convert an untrustworthy prediction or model into a reliable one. To tackle this challenge, the authors introduce LIME, a novel explanation technique designed to elucidate the predictions of any classifier in an interpretable and faithful manner. This is achieved by constructing an interpretable model locally around each prediction. Additionally, they propose a method for explaining models by showcasing representative individual predictions along with their explanations in a non-redundant manner, framing this task as a submodular optimization problem. The flexibility and applicability of these methods are demonstrated through experiments involving different types of models for text (such as random forests) and image classification (like neural networks). The utility of explanations is showcased through various scenarios that necessitate trust, including determining whether to trust a prediction, selecting between models, enhancing an unreliable classifier, and pinpointing reasons why a classifier should not be trusted. Overall, this paper provides valuable insights and practical techniques for interpreting machine learning models' predictions in order to foster trust and enable informed decision-making in diverse applications.