ELI5 works in Python 2.7 and Python 3.4+. Currently it requires scikit-learn 0.18+. You can install ELI5 using pip:
pip install eli5
ELI5 is a Python package which helps to debug machine learning classifiers and explain their predictions. It provides support for the following machine learning frameworks and packages:
scikit-learn. Currently ELI5 allows to explain weights and predictions of scikit-learn linear classifiers and regressors, print decision trees as text or as SVG, show feature importances and explain predictions of decision trees and tree-based ensembles.
Pipeline and FeatureUnion are supported.
ELI5 understands text processing utilities from scikit-learn and can highlight text data accordingly. It also allows to debug scikit-learn pipelines which contain HashingVectorizer, by undoing hashing.
XGBoost - show feature importances and explain predictions of XGBClassifier, XGBRegressor and xgboost.Booster.
LightGBM - show feature importances and explain predictions of LGBMClassifier and LGBMRegressor.
lightning - explain weights and predictions of lightning classifiers and regressors.
sklearn-crfsuite. ELI5 allows to check weights of sklearn_crfsuite.CRF models.
ELI5 also provides TextExplainer which allows to explain predictions of any text classifier using LIME algorithm (Ribeiro et al., 2016). There are utilities for using LIME with non-text data and arbitrary black-box classifiers as well, but this feature is currently experimental.
Explanation and formatting are separated; you can get text-based explanation to display in console, HTML version embeddable in an IPython notebook or web dashboards, JSON version which allows to implement custom rendering and formatting on a client, and convert explanations to pandas DataFrame objects.
There are two main ways to look at a classification or a regression model:
- inspect model parameters and try to figure out how the model works globally;
- inspect an individual prediction of a model, try to figure out why the model makes the decision it makes.
If the ML library you’re working with is supported then you usually can enter something like this in the IPython Notebook:
import eli5 eli5.show_weights(clf)
and get an explanation like this:
Depending on an estimator, you may need to pass additional parameters to get readable results - e.g. a vectorizer used to prepare features for a classifier, or a list of feature names.
Supported arguments and the exact way the classifier is visualized depends on a library.
To explain an individual prediction (2) use
function. Exact parameters depend on a classifier and on input data kind
(text, tabular, images). For example, you may get text highlighted like this
if you’re using one of the scikit-learn vectorizers with char ngrams:
For some of classifiers inspection and debugging is easy, for others this is hard. It is not a rocket science to take coefficients of a linear classifier, relate them to feature names and show in an HTML table. ELI5 aims to handle not only simple cases, but even for simple cases having a unified API for inspection has a value:
- you can call a ready-made function from ELI5 and get a nicely formatted result immediately;
- formatting code can be reused between machine learning frameworks;
- ‘drill down’ code like feature filtering or text highlighting can be reused;
- there are lots of gotchas and small differences which ELI5 takes care of;
- algorithms like LIME (paper) try to explain a black-box classifier through a locally-fit simple, interpretable classifier. It means that with each additional supported “simple” classifier/regressor algorithms like LIME are getting more options automatically.
In ELI5 “explanation” is separated from output format:
Explanation instances; then functions from
eli5.formatters can be used to get HTML, text, dict/JSON
or pandas DataFrame representation of the explanation.
It is not convenient to do that all when working interactively in IPython
notebooks, so there are
eli5.show_prediction() functions which do explanation and formatting
in a single step.
Explain functions are not doing any work by themselves; they call
a concrete implementation based on estimator type.
sklearn.linear_model.LogisticRegression classifier is passed
as an estimator.