Anchor Explanations

Anchor Explanations

Anchor Explanations is a powerful interpretability technique used in machine learning (ML) to provide human-understandable explanations for predictions made by complex models. This method is particularly useful for black-box models, where the internal workings are not easily interpretable.


Anchor Explanations is a model-agnostic method that explains the behavior of machine learning models by identifying sets of ‘anchors’. These anchors are conditions that sufficiently guarantee a prediction, regardless of the values of other features. The technique is based on the concept of coverage (how much the anchor applies) and precision (how much the anchor guarantees the prediction).

How it Works

Anchor Explanations operates by finding an ‘anchor point’ in the input space that, when present, ensures the model’s prediction with a high degree of certainty. It starts by identifying a point in the feature space, then iteratively adds other features to this point until it finds a minimal set of features (the anchor) that, when fixed, will almost always lead to the same prediction.

The algorithm uses a combination of local search methods and reinforcement learning to find these anchors. It starts with a small candidate set of features and expands it by adding or removing features based on their impact on the prediction.

Use Cases

Anchor Explanations are used in various domains where interpretability is crucial. They are particularly useful in healthcare, finance, and legal sectors where understanding the reasoning behind a prediction is as important as the prediction itself. For instance, in healthcare, a doctor might need to understand why a particular ML model predicted a certain disease for a patient.


  1. Model Agnostic: Anchor Explanations can be applied to any machine learning model, making it a versatile tool for interpretability.

  2. High Precision: The anchors provide high precision explanations, meaning they are reliable and can be trusted to maintain the same prediction when they apply.

  3. Human Understandable: The explanations provided by anchors are simple and understandable, making them useful for non-technical stakeholders.


  1. Computationally Intensive: Finding the right set of features to act as an anchor can be computationally expensive, especially for high-dimensional data.

  2. Local Explanations: Anchor Explanations provide local interpretability, meaning they explain individual predictions rather than the overall behavior of the model.

  • Interpretability in Machine Learning: The degree to which a human can understand the cause of a decision made by a machine learning model.

  • Model Agnostic Methods: Interpretability methods that can be applied to any machine learning model.

  • Black-Box Models: Machine learning models whose internal workings are not easily interpretable or understandable.


  1. Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. “Anchors: High-Precision Model-Agnostic Explanations.” AAAI Conference on Artificial Intelligence. 2018.

  2. Molnar, Christoph. “Interpretable Machine Learning.”, 2019.

Last Updated: August 14, 2023