This article is a continuation of my series of articles on Model Interpretability and Explainable Artificial Intelligence. If you haven’t read the first two articles I would highly recommend you to do so.
The first article of the series, ‘Introduction to Machine Learning Model Interpretation’, covers the basics of Model Interpretation. The second article, ‘Hands-on Global Model Interpretation’, goes over the details of global model interpretation and how to apply it to a real-world problem using Python.
In this article, we will pick up where we left off by diving into local model interpretation. First we will take a look at what local model interpretation is and what it can be used for. Then we will dive into the theory of two of it’s most popular methods, Lime (Local Interpretable Model-Agnostic Explanations) and Shapley Value, and apply them to get information about individual predictions on the heart disease data-set.
What is local model interpretation?
Local model interpretation is a set of techniques aimed at answering questions like:
- Why did the model make this specific prediction?
- What effect did this specific feature value have on the prediction?
Using our domain expertise combined with the knowledge gained using local model interpretation we can make decisions about our model like if it is right for our problem or if it is not functioning as it’s supposed to.
The use of local model interpretation is getting increasingly important these days because a company (or individual) must be able to explain the decisions of their model. Especially if it is getting used in a domain that requires a lot of trust like for medicine or finances.
LIME (Local Interpretable Model-Agnostic Explanations)
Despite widespread adoption, machine learning models remain mostly black boxes. Understanding the reasons behind predictions is, however, quite important in assessing trust, which is fundamental if one plans to take action based on a prediction, or when choosing whether to deploy a new model. — Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. “Why should I trust you?”
Figure 2: Lime Example
Concept and Theory
Lime, Local Interpretable Model-Agnostic, is a local model interpretation technique using Local surrogate models to approximate the predictions of the underlying black-box model.
Local surrogate models are interpretable models like Linear Regression or a Decision Trees that are used to explain individual predictions of a black-box model.
Lime trains a surrogate model by generating a new data-set out of the datapoint of interest. The way it generates the data-set varies dependent on the type of data. Currently Lime supports text, image and tabular data.
For text and image data LIME generates the data-set by randomly turning single words or pixels on or off. In the case of tabular data, LIME creates new samples by permuting each feature individually.
The model learned by LIME should be a good local approximation of the black box model but that doesn’t mean that it’s a good global approximation.
LIME is a good pick for analyzing predictions because it can be used for any black-box model no matter if it is a deep neural network or an SVM. Furthermore LIME is one of the only interpretation techniques that works for tabular, text and image data.
The definition of neighborhood of the data-point of interest is very vague which is a huge problem because that means that you have to try different kernel settings for most applications.
Furthermore, the explanations can be unstable meaning that the explanation of to very close data-points can vary greatly.
Example and Interpretation
LIME can be used in Python with the Lime and Skater packages which make it really easy to use LIME with models from popular machine learning libraries like Scikit Learn or XGBoost. For this article, I’ll work on the Heart Disease data-set, a simple classification data-set.
To analyze the predictions of a tabular data-set like the Heart Disease data-set we need to create a LimeTabularExplainer.
import lime import lime.lime_tabular rf = RandomForestClassifier().fit(X_train, y_train) explainer = lime.lime_tabular.LimeTabularExplainer(np.array(X_train), feature_names=df.drop(['target'], axis=1).columns.values, class_names=rf.classes_, discretize_continuous=True)
Now we can use the explain_instace method to explain a single prediction.
exp = explainer.explain_instance(X_test.iloc, rf.predict_proba) exp.show_in_notebook(show_table=True, show_all=False)
Figure 3: Lime Output
Here we can nicely see what feature values influenced the prediction. For this data-point, the cp and oldpeak values had the most negative influence on the prediction.
For more examples on how to use Lime check out the “Tutorials and API” section in the README file.
The Shapley value is a method for assigning payouts to players depending on their contribution to the total payout. But what does this have to do with Machine Learning?
In the case of machine learning the “game” is the prediction task for a data point. The “gain” is the prediction minus the average prediction of all instances and the “players” are the feature values of the data point.
The Shapley value is the average marginal contribution of a feature value across all possible coalitions. It is an excellent way of interpreting an individual prediction because it doesn’t only give us the contribution of each feature value but the scale of these contributions is also right which isn’t the case for other techniques like LIME.
When using Shapley values the prediction and the average prediction is fairly distributed among the feature values of the instance. This makes Shapley Values perfect for situations that need a exact explanation. Furthermore Shapley value ist the only local explanation method with solid theory behind it.
The Shapley value requires a lot of computing time. Most often, only the approximate solution is feasible. Another disadvantage is that you need access to the data to calculate the Shapley value. This is because you need the data to replace parts of the instance of interest. This problem can only be avoided if you can create data instances that look like real data instances.
Example and Interpretation
The shap package can be used for multiple kinds of models like trees or deep neural networks as well as different kinds of data including tabular data and image data.
To use shap with the heart disease data-set we can create a TreeExplainer which can explain the predictions of tree models like a Random Forest and then use the shap_values and force_plot methods to analyze predictions.
import shap shap.initjs() model_lgb = lgb.LGBMClassifier().fit(X_train, y_train) # Create a object that can calculate shap values explainer = shap.TreeExplainer(model_lgb) # Calculate Shap values shap_values = explainer.shap_values(np.array(X_test)) shap.force_plot(explainer.expected_value, shap_values[0,:], X_test.iloc[0,:])
In the image, we can see how each feature value affects the prediction. In this case, the predicted value is 0 which means the patient doesn’t have a heart disease. The most significant feature values are cp=0 and ca=3. This is indicated by the length of the arrows.
Local model interpretation is a set of techniques trying to help us understand individual predictions of machine learning models. LIME and Shapley value are two techniques to understand individual predictions.
Lime is very good for getting a quick look at what the model is doing but has problems with consistency. Shapley value delivers a full explanation which makes it a lot more exact then Lime.
That’s all from this article. If you have any questions or just want to chat with me feel free to leave a comment below or contact me on social media. If you want to get continuous updates about my blog make sure to follow me on Medium and join my newsletter.