naive bayes
fully connected bayes net
Area(full bayes) > Area(naive bayes)
best -> which attribute should we split on (according to the heuristic)
tree -> build a new tree on that attribute best
left is better! -> less uncertainty -> less entropy!
efficient way for coding
entropy -> the amount of uncertainty of our distribution
should split on cylinder
unexpandable means all the samples here are stochastic
go all the way to the bottom -> overfit
-> walk back and prune
why go all the way to the bottom?
because sometimes we need multiple features to determine the results
eg. XOR
test set error reduced by pruning: 21% -> 16%
网友评论