Sifting the Sea: Finding Just Enough to Predict from Too Much

Abstract

Every time Netflix suggests what to watch or your phone predicts your next word, it is "predicting" based on a classification or regression model built using huge amounts of data. But here is the catch -- using all that data can be slow, messy, and even unnecessary. What if we could make smart predictions by using just the right amount of data? In this talk, we will explore how picking a small, well-chosen part of a dataset -- instead of the entire big dataset -- can still lead to accurate results. Through simple ideas and visual examples, we’ll see how this approach, called "subdata selection", can help us learn faster and smarter from the data around us.