Interpretation

Interpretation of a model needs to make meaningful sense. First of all, one has to consider that an interpretation must not be confusing. For that reason, a good practice is always to consider a lower dimension model (if possible), by considering the features that have the highest impact on the results. For that reason, we need to do the feature importance and feature selection. This means essentially to find the features that can capture the most information in the data. We will introduce few methods in the future.

Second, there are ML methods that by their nature are already interpretable. These methods are very helpful, and if a good fit, can be used directly to interpret the results of an ML model. For instance, trees (including classifiers and regressions), fuzzy classifiers, linear regression, logistic regression, etc, are interpretable. However, we always need to be careful not to use these methods in a way that would not really help for interpretation. That is why one always needs to consider lower feature dimensions (feature selection), as discussed above.

Third, recently there have been methods introduced that can be used to interpret any ML model regardless of their complexity (or un-interpretability). These approaches are known as model agnostic interpretation methods.

Low dimension interpretation
Interpretable models (trees, fuzzy classifiers, etc)
Model agnostic interpreters (like LIME)

Low dimension

Here we consider data from insurance that the clients are labeled with two color light blue and light brown, for highly scored client and low scored client. After doing the feature selection we found that two features are more important, age (y-axis) and making previous claims (x-axis). As one can see the labeled data can properly be distinguished based on the values of the two features. This is very helpful, since we can then have better insight if the data with these two dimensions are really well interpretable.

Interpretable algorithms

For instance, we can use trees to find a good interpretation. However, one needs to be careful that not to make the interpretation so confusing that making legitimate propositions become very challenging. That is why except for limiting the dimension, we also have to consider a shallow tree and limit the depth of the tree.

Model agnostic interpreter

Now let us consider other models, to fit into the data. We consider the following models:

•Logistic Regression

•Linear Discriminant Analysis

•K-neighbors Classifier

•Gaussian NB

•Decision Tree Classifier

•Random Forest Classifier

•Gradient Boosting Classifier

•Ensemble Vote Classifier

As it is clear, there are models that are a very good fit, especially the last four models, among which only the decision tree is interpretable. That is why it would be very helpful if we can find a way to interpret all models regardless of the fact that they are interpretable or not.

LIME, a local interpretation

Here we present a model agnostic interpretation method that essentially does not care about the type of model that is fit. The methods is called LIME (Local Interpretable Model-agnostic Explanations). This particular way of interpretation is local, and that is why we only can provide interpretation for any individual. The meaning of this is that the model will provide the reasons that why a new individual sample can be classified with the label that the model proposes. The method essentially need first a good model fit to the data. Then for a new sample, one need to perturbate the other existing samples around that and create a set of (fake) data. Then one will use an interpretable model, like linear regression or a tree, to fit the created-data.

The method is local but this approach is not surprising. A good example is when an individual person visits a doctor. A doctor essentially can be regarded as a classifier to say if the individual is seeking or well. However, the reasons that why an individual is seeking or not is to some extend specifically to the person.

Slides

Interpret-site.pdf

Page updated

Google Sites

Report abuse