How does tree bagger handle NaN values

In building a random forest classifier I have some features with a large amount of NaN values, but it is not clear to me how Tree Bagger handles these NaNs. I've seen quite a bit of documentation of how that is handled in other high level programming languages, but I don't see explicitly how this is done in Matlab. Can anyone point me in the right direction so I can understand the default settings for this or user specified settings?

Answer 1

[Puru Kathuria]

General rules that are followed while NaN or missing values are encountered:

• Rule1: The algorithm simply discards the data points where all the features have NaN values and does not use them while training.

• Rule 2: If a data point have a few NaN feature values then the algorithm will find the split on the basis of valid values first.