|
I understand that one way to avoid the probability estimation going to zero is to do smoothing. I have found literature like this link http://www.cs.cmu.edu/~tom/mlbook/NBayesLogReg.pdf that asks me to correct by adding "number of values the attribute takes" in the denominator and 1 in the numerator. (Laplace Smoothing.) while this makes sense in the case of discrete valued attributes, how do i go about doing this in the case where my attributes have continuous values? i'm assuming a normal distribution on the attributes. |
|
In Gaussian Naive Bayes, the zero division problem doesn't occur. You can use it unsmoothed without problems. |
