What was true in trying to understand our data is true for trying to understand our classifier: visualization is the first step in understanding a system. We know the SVM somehow came up with a decision boundary that allowed us to correctly classify 80 percent of the test samples. But how can we find out what that decision boundary actually looks like?
For this, we will borrow a trick from the guys behind scikit-learn. The idea is to generate a fine grid of x and y coordinates and run that through the SVM's predict method. This will allow us to know, for every (x, y) point, what target label the classifier would have predicted.
We will do this in a dedicated function, which we call plot_decision_boundary ...
