Error measure and Hoeffding inequality

[v_venky]

In the 1st chapter when Hoeffding is used, the error measure is simple mismatch -- there is no penalty associated with the different flavors of mismatch as in error measures. Is there a version of Hoeffding which we could use with error measure as well? If so how does it look? I imagine the red balls would now have weights of some sort -- two types of red balls for the two kinds of errors in a binary classifier.

[htlin]

Yes, putting in the weights is certainly possible. This lecture

https://www.csie.ntu.edu.tw/~htlin/m...08_handout.pdf

contains some discussions. It was taught in Mandarin (on Youtube) though.

Hope this helps.