Including Thermometer Encoding method

[mahedeeb]

Thermometer encoding is like one-hot encoding, but it represents magnitude instead of a categorical variable. So for example, if you need to represent a magnitude of 7 in a vector of 10 elements, then the thermometer encoding will be [0,0,0,1,1,1,1,1,1,1]. In short, its an alternative way to encode magnitude. It does seem interesting, but one’s intuition would be that it wouldn’t make much of a difference versus real numbers. In fact, it should be detrimental in that it requires more storage.

For more information, please check the following paper:
https://openreview.net/pdf?id=S18Su--CW

another source:
https://medium.com/intuitionmachine/the-shocking-effectiveness-of-thermometer-encoding-f155ed9c581b

you can check this kernel in Kaggle:
https://www.kaggle.com/superant/oh-my-cat