can reliefF() function deal with NaNs in my matrix?

Hello,

I have an (n*m double) matrix, where n (rows) is the number of my samples and m (columns) is the number of my features, which are all discrete (i.e. categorical). The mth column in the data represents my binary outcome. I have tried using relieff() function to return the importance of my predictor features based on my outcome feature.

This is what my data looks like (let's say for simplicity's sake: I have 4 predictor and 1 outcome feature for 3 samples):

    matrixdata = [1, 2, 3, NaN, 2; 5, 1, NaN, 2, 1; NaN, 3, NaN, 2, 1];

This is how I call the relieff() on my data:

    X = matrixdata(:,1:(end-1));
    Ylogical = matrixdata(:,end)== 1;
    [ranked,weights] = relieff(X,Ylogical,10, 'categoricalx', 'on');

In this case, does relieff() disregard the NaNs in the data or does it treat NaNs as a separate category of that predictor feature column? Obviously, the former is what I would prefer.

Many thanks, Berkan

Answer 1

[Wayne King]

Hi Berkan, relieff() removes NaNs in both your predictor and response variables. So your preference is the way it is implemented.

[Berkan Sesen]

Hi Wayne, thank you very much, it is a relief!

Best,

berkan

Answer 2

[Berkan Sesen]

Hi, apparently the reliefF function has a sub function removeNaNs(X,Y) as below:

function [X,Y] = removeNaNs(X,Y)
% Remove observations with missing data
NaNidx = bsxfun(@or,isnan(Y),any(isnan(X),2));
X(NaNidx,:) = [];
Y(NaNidx,:) = [];

This effectively gets rid of all rows that contain any NaNs and is not ideal, especially for my case, since it leaves me with X=[] and Y=[] (i.e. no observations!)

How can I tackle this? Would replacing all NaN's with a random category, e.g. 99999, help? By doing this, I am introducing a new node state for all the predictor features so I guess it is not ideal.

Thanks, Berkan