You can apply a cost matrix to force the classifier to make fewer mistakes on a subset of your classes, even if they aren't very frequent. So in a two class problem, you can prioritize the accuracy of the first class by using a matrix like C = [0,10;1,0] which penalizes misclassifications of the first class by a factor of 10, and then pass that cost matrix to the training routine using the 'Cost' parameter, like
C = [0,10;1,1];
model = fitc...
Right now, you can't do this from within the ClassificationLearner app, you have to step back to the programmatic workflow (but you can export code for the model you were exploring in the app with the "Export Model" button).
NaNs are ignored in the data by default.