>>11368795>>11368780Imagine you have a dataset of people, with their height, and their age (in the interval 0-25).
0,30
1,70 (sorry i dont know how tall babies are
2,90
3,100
4,110
this is the form xi, yi, where you could use x (their age) to predict their height yi
Usually you would shuffle it so that each i aka each (xi, yi) gets a new order. That way you don't get any bias in the dataset, since we know each sample is supposed to be independent. This would give
3,100
0,30
2,90
1,70
4,110
What this guy is saying his manager says gets better results is splitting it into two data sets. X and Y, then suffling the datasets independently, before merging them back together. This destroys the data completely, since the new (xi, yi) no longer would correspond to a person. There would no longer be any relationship between the xis and the yis.
4,100
1,90
2,30
0,110
3,70
If you instead of shuffling the data, sort it, you would do the same, aka breaking any relationship between xes and ys, but now the biggest xes would automatically be put together with the biggest ys, and you force a positive relationship between the data, no matter the starting data.
