Top Machine Learning Model Interpretation Tools

As important as making machine learning models is, the ability to interpret and explain the decisions made by an ML model is crucial for ensuring the fairness, accountability, and trustworthiness of the predictions. Not to mention, it also helps make it look less like a black box.

Being able to interpret the model allows developers to move from the question of ‘what did the model predict?’ to ‘how did the model predict this?’ To achieve this, a lot of techniques have been developed over the years that vary slightly from each other and are used for different models by way of different approaches.

Here are some of the top tools and techniques that developers use for interpreting machine learning models!

ELI5

An acronym for ‘Explain like I’m 5’, ELI5 is a Python library and is fairly easy to use for explaining and interpreting ML models. The technique interprets models in a local/global scope way, rather than specific/agnostic way. Currently, it requires scikit-learn 0.18+ to run.

The library allows explanation of weights and prediction of linear classifiers and regressors, and prints decision trees as text or SVGs. It also shows feature importances along with explaining predictions of decision trees and tree-based ensembles.

Download our Mobile App

The primary usage of this is to inspect model parameters and figuring out how the model works globally, or inspecting individual predictions of the model and relating it to the decision making process of the model. The model is currently only available for Keras, XGBoost, LightGBM, CatBoost, lightning, and sklearn-crfsuite.

SHAP 

Abbreviation for SHapley Additive exPlanations, SHAP takes a different approach to explain the output of any ML model using game-theory. It explains the prediction of an instance through computation of the contribution of each feature relative to that prediction using Shapley values.

Shapley values is a method derived from coalitional game theory and iterates how the payout can be distributed among the features. These values are the average marginal contribution of every single feature value across all possible coalitions. It is considered to be the most intuitive approach for interpretability of models though it can often be complicated and overwhelming for beginners. 

SHAP explanation is expressed mathematically as: 

LIME

Local Interpretable Model-agnostic Explanations (LIME) which means that the framework analyses individual predictions and gives extensive insights about the predictions of any ML classifier. The repository is mostly used for Javascript and partially by Python developers. 

Being model agnostic, LIME treats every model as a black box. This means that it makes interpretations outside the model and thus works on a wide range of models. LIME offers three disparate methods for interpretation that deal with different kinds of data: 

• Text Interpretation
• Tabular Interpretation
• Image Interpretation

Anchors

Anchors is similar to SHAP and LIME for generating local explanations, but instead of interpretable models, this tool expresses explanations as simple and easy to understand ‘If–then’ rules. As the name suggests, rule anchors the prediction so that the changes in other values do not affect the prediction.

This technique is ideally used for image classification models as it identifies the minimal set of image regions, denoting them as anchors, that allows it to make decisions. These anchors are created by identifying regions that have high class activation values, meaning the area where the model is making the prediction.

One key advantage of the anchors technique is that it allows for the generation of human-interpretable explanations, which can be useful for understanding how a model is making its decisions and identifying potential sources of bias.

MLxtend

Otherwise known as Machine Learning Extensions, MLxtend is a Python library for data analysis. The library offers features such as:

• PCA correlation circle: The library lets you look at the results by plotting a PCA correlation circle and then correlate the features and the Principal Component. These are then plotted as vectors on a unit circle with Principal components as the axes. These axes show the variance for the specific Principal Components.

• Bias–Variance decomposition: To tackle the bias–variance tradeoff, which is avoiding underfitting and overfitting in a model, MLxtend gets a bias–variance score for predictive models and decomposes the generalisation error into bias, variance, and error score.

• Plotting decision boundaries and regions of the model: MLxtend also allows looking at the decision boundary of the model in two dimensions to figure out how the model is differentiating among data points in different classes. Though this can only be done for two features at the time of visualisation.

Conclusion 

As models become increasingly complex, interpreting them is the most important aspect for deploying them. Further techniques and tools are being developed for it and the tools covered here are not the only ones available, but are among the ones used most prominently.