Unit 4.3 Zero-rule baseline (majority classifier)

[harishb00]

Below image is a snapshot from MNIST MLP Notebook

@rasbt
Shouldn't the majority_class be from training dataset? Shouldn't it be majority_class = train_counter.most_common(1)[0]?
Because from the training set we look for the majority class and predict the same for the real world unseen data.

[rasbt]

Good point. In a practical setting where you want to implement a MajorityClass classifier, I'd say you are absolutely right. Here, I was more focused on seeing the best possible test accuracy without actually using a model. The majority class derived from the training set would always be equal or lower than the one derived from a test set, so it's more like the "best lucky guess of a majority class classifier" here.