BinarizeMatrix
binarizeMatrix is a Matlab function that takes as an input one or more normalized time series matrices, Z, and returns up to two transformations of these matrices, BIN and SCALED. A Z-score threshold, thresh is applied to the normalized values: For all values greater than +thresh are set to 1; all values less than -thresh are set to 0. For the SCALED transformation, the intermediate values are scaled so that the columnar mean of the values between 0 and 1 is 0.5. For the BIN transformation, all fractional values are set to NaN. Additionally, a clipping value, clip can be supplied: all values with an absolute value exceeding the clip value are set to NaN. If a threshold is not supplied, 0 is used by default (i.e., all positive values are set to 1, and negative values are set to 0). If a clipping value is not used, no outlier values are filtered.
Defensible Z threshold values
AUC | Z |
---|---|
.66 | .44 |
.75 | .68 |
.80 | .84 |
.90 | 1.28 |
.95 | 1.65 |
.99 | 2.33 |
Function Source Code
function [BIN, SCALED]=binarizeMatrix(Z, varargin) %function [BIN, SCALED]=binarizeMatrix(Z, varargin) %binarizeMatrix takes a normalized matrix, Z, and scales and binarizes its %elements: %positive values are set to 1 and negative values ar set to 0. If thresh is %set to a nonzero value, elements with an absolute value less than thresh %are set to NaN. If a clipping value, clip, is provided, elements with an %absolute value exceeding clip are first set to NaN before binarizing. %Required Parameters: % Z: a normalized matrix or cell array % %Optional Parameters: % thresh: the z-score threshold above which the absolute values of the % normalized values in Z will be set to 1 (for positive values above the % threshold) or 0 (for negative values below the negative threshold) % clip: the z-score above which the absolute values will be set to NaN % precision: values are rounded to PRECISION decimal points %Sample usage: %thresh=1.96; %clip=3.0; %[BIN, SCALED]=binarizeMatrix(Z, 'thresh', thresh, 'clip', clip, 'precision', 2); % options = struct('thresh',1, ... 'clip', inf, ... 'precision', 1); % read the acceptable names optionNames = fieldnames(options); % count arguments nArgs = length(varargin); if (isstruct(varargin)) options=varargin; else if round(nArgs/2)~=nArgs/2 error('This function needs propertyName/propertyValue pairs') end for pair = reshape(varargin,2,[]) % pair is {propName;propValue} inpName = lower(pair{1}); % make case insensitive if any(strcmp(inpName,optionNames)) % overwrite defaults options.(inpName) = pair{2}; else error('%s is not a recognized parameter name',inpName) end end end if(options.thresh==0) options.thresh=realmin; %workaround to allow a threshold of approx. zero end if (~iscell(Z)) %Z is not a cell array, which is the easiest case SCALED=scaleMatrix(Z,options.clip, options.thresh, options.precision); BIN=SCALED; BIN(0<BIN & BIN<1)=nan; else if(iscell(Z)) %M is a cell array, so we have to apply droprows and zscore to each %cell element SCALED=cellfun(@(x) scaleMatrix(x, options.clip, options.thresh, options.precision), Z, 'UniformOutput', false); BIN=cellfun(@(x) nanBin(x), SCALED, 'UniformOutput', false); end end
%%Nested utility function scale matrix function SM=scaleMatrix(M, clip, thresh, precision) M(abs(M)>clip)=nan; %remove items to be clipped M=M/thresh;%scale in terms of thresholds M(M>1)=1;%anything above +threshold=1 M(M<-1)=-1; %anything below -threshold=-1 M=M+1;%shift values from range -1:1 to 0:2 M=M/2;%now values range from 0:1 %subtly shift values so that mean value for each column is ~0.5 meansignal=nanmean(M);%what is the current mean of each column? %use log to determine exponent required to shift each mean to 0.5 centerexp=repmat(power(log10(meansignal),-1).*log10(0.5),size(M,1),1); %raise each value to the required exponent. 0s and 1s will be %unaffected precis=power(10,precision); SM=round(precis*M.^centerexp)/precis; end %%Nested utility function sets all non 0-1 values to nan function BIN=nanBin(M) BIN=M; BIN(0<BIN & BIN<1)=nan; end end