[wenc]

> 'too computationally expensive for machine learning/big data world'

But not for normal-sized datasets...

I've noticed a shift in the past year where (new?) people are thinking ML is synonymous wth NN/Deep Learning models.

To me ML has always encompassed statistical learning techniques, most of which work very well on normal-sized datasets (thousand to a million rows, definitions differ). Most classical/statistical methods work just fine at this scale, including the method in the linked article.

Lately I've also been thinking: given certain patterns of regularity in data, and outside of certain domains involving images/sound/language, we don't really need large-scale datasets to train our models. Good models can be trained on carefully selected samples without significant loss of fidelity, which opens up the scope of the types of models that can be deployed.