fcol: Generic colour module for forestFloor objects
In forestFloor: Visualizes Random Forests with Feature Contributions

Description Usage Arguments Details Value Author(s) Examples

This colour module colour observations by selected variables. PCA decomposes a selection more than three variables. Space can be inflated by random forest variable importance, to focus coloring on influential variables. Outliers(>3std.dev) are automatically suppressed. Any colouring can be modified.

fcol(ff, cols = NULL, orderByImportance = NULL, plotTest=NULL, X.matrix = TRUE, 
     hue = NULL, saturation = NULL, brightness = NULL,
     hue.range  = NULL, sat.range  = NULL, bri.range  = NULL,
     alpha = NULL, RGB = NULL, byResiduals=FALSE, max.df=3,
     imp.weight = NULL, imp.exp = 1,outlier.lim = 3,RGB.exp=NULL)

`ff`	a object of class "forestFloor_regression" or "forestFloor_multiClass" or a matrix or a data.frame. No missing values. X.matrix must be set TRUE for "forestFloor_multiClass" as colouring by multiClass feature contributions is not supported.
`cols`	vector of indices of columns to colour by, will refer to ff$X if X.matrix=T and else ff$FCmatrix. If ff itself is a matrix or data.frame, indices will refer to these columns. If NULL, then all cols are selected.
`orderByImportance`	logical, should cols refer to X column order or columns sorted by variable importance. Input must be of forestFloor -class to use this. Set to FALSE if no importance sorting is wanted. Otherwise leave as is. If NULL, then TRUE for forestFloor objects and FALSE for data.frames and matrices.
`plotTest`	NULL(plot by test set if available), TRUE(plot by test set), FALSE(plot by train), "andTrain"(plot by both test and train)
`X.matrix`	logical, TRUE will use feature matrix, FALSE will use feature contribution matrix. Only relevant if ff input is forestFloor object.
`hue`	value within [0,1], hue=1 will be exactly as hue = 0 colour wheel settings, will skew the colour of all observations without changing the contrast between any two given observations. NULL is auto.
`saturation`	value within [0,1], mean saturation of colours, 0 is grey tone and 1 is maximal colourful. NULL is auto.
`brightness`	value within [0,1], mean brightness of colours, 0 is black and 1 is lightly colours. NULL is auto.
`hue.range`	value within [0,1], ratio of colour wheel, small value is small slice of colour wheel those little variation in colours. 1 is any possible colour except for RGB colour system. NULL is auto.
`sat.range`	value within [0,1], for colouring of 2 or more variables, a range of saturation is needed to obtain more degrees of freedom in the colour system. But as saturation of is preferred to be >.75 the range of saturation cannot here exceed .5. If NULL sat.range will set widest possible without exceeding range.
`bri.range`	value within [0,1], for colouring of 3 or more variables, a range of brightness is needed to obtain more degrees of freedom in the colour system. But as brightness of is preferred to be >.75 the range of saturation cannot here exceed .5. If NULL bri.range will set widest possible without exceeding range.
`alpha`	value within [0;1] transparency of colours. NULL is auto. The more points the less alpha, to avoid overplotting.
`RGB`	logical TRUE/FALSE, RGB=NULL: will turn TRUE if one and only one variable is selected by cols input RGB=TRUE: "Red-Green-Blue": A non-linaer mapping from data to hsv. The gradient span from red to green to blue. Can still be altered by hue, saturation, brightness etc. RGB=FALSE: "True-colour-system": Linear mapping from data to hsv colours. Sometimes more confusing, due to low contrast and over/under saturation.
`byResiduals`	logical, should colour gradient span the residuals of main effect fit(overrides X.matrix=). If no fit has been computed "is.null(ff$FCfit)", a temporarily main effect fit will be computed. Use ff = convolute_ff(ff) to only compute once and/or to modify fit parameters.
`max.df`	integer 1, 2, or 3 only. Only for true-colour-system, the maximal allowed degrees of freedom (dimensionality) in a colour gradient. If more variables selected than max.df, PCA decomposes to requested degrees of freedom. max.df = 1 will give more simple colour gradients.
`imp.weight`	Logical?, Should importance from a forestFloor object be used to weight selected variables? Not possible if input ff is a matrix or data.frame. If randomForest(importance=TRUE) during training, variable importance will be used. Otherwise the more unreliable gini_importance coefficient.
`imp.exp`	exponent to modify influence of imp.weight. 0 is no influence. -1 is counter influence. 1 is linear influence. .5 is square root influence etc..
`outlier.lim`	number from 0 to Inf. Any observation which univariately exceed this limit will be suppressed, as if it actually where on this limit. Normal limit is 3 standard deviations. Extreme outliers can otherwise reserve alone a very large part of a given linear colour gradient. This leads to visualization where outlier have one colour and any other observation have one single other colour.
`RGB.exp`	value between ]1;>1]. Defines steepness of the gradient of the RGB colour system Close to one green middle area is missing. For values higher than 2, green area is dominating

fcol() is designed for the lazy programmer who likes to visualize a data set and/or a model-mapping swiftly without typing to much, check the examples for the typical use. To extra visual dimensions with colour gradients is useful to discover multivariate interactions in your randomForest model. fcol() both works for forestFloor, but also for any data.frame and matrix in general. Whenever an input parameter is not set or set to NULL, fcol() should choose a reasonble auto setting depending on possible other settings. In general if one manually set a parameter to something else than NULL it will override any possible auto setting.

a character vector specifying the colour of any observations. Each elements is something like "#F1A24340", where F1 is the hexadecimal of the red colour, then A2 is the green, then 43 is blue and 40 is transparency.

Soren Havelund Welling

## Not run: 
#example 1 - fcol used on data.frame or matrix
library(forestFloor)
X = data.frame(matrix(rnorm(1000),nrow=1000,ncol=4))
X[] = lapply(X,jitter,amount = 1.5)

#single variable gradient by X1 (Unique colour system)
plot(X,col=fcol(X,1))
#double variable gradient by X1 and X2 (linear colour system)
plot(X,col=fcol(X,1:2))
#triple variable gradient (PCA-decomposed, linear colour system)
plot(X,col=fcol(X,1:3))
#higher based gradient    (PCA-decomposed, linear colour system)
plot(X,col=fcol(X,1:4))


#force linear col + modify colour wheel
plot(X,col=fcol(X,
                cols=1, #colouring by one variable
                RGB=FALSE,
                hue.range = 4, #cannot exceed 1, if colouing by more than one var
                               #except if max.df=1 (limits to 1D gradient)
                saturation=1,
                brightness = 0.6))

#colour by one dimensional gradient first PC of multiple variables
plot(X,col=fcol(X,
                cols=1:2, #colouring by multiple
                RGB=TRUE, #possible because max.df=1
                max.df = 1, #only 1D gradient (only first principal component)
                hue.range = 2, #can exceed 1, because max.df=1
                saturation=.95,
                brightness = 0.8))

##example 2 - fcol used with forestFloor objects
library(forestFloor)
library(randomForest)

X = data.frame(replicate(6,rnorm(1000)))
y = with(X,.3*X1^2+sin(X2*pi)+X3*X4)
rf = randomForest(X,y,keep.inbag = TRUE,sampsize = 400)
ff = forestFloor(rf,X)

#colour by most important variable
plot(ff,col=fcol(ff,1))

#colour by first variable in data set
plot(ff,col=fcol(ff,1,orderByImportance = FALSE),orderByImportance = FALSE)

#colour by feature contributions
plot(ff,col=fcol(ff,1:2,order=FALSE,X.matrix = FALSE,saturation=.95))

#colour by residuals
plot(ff,col=fcol(ff,3,orderByImportance = FALSE,byResiduals = TRUE))

#colour by all features (most useful for colinear variables)
plot(ff,col=fcol(ff,1:6))

#disable importance weighting of colour
#(important colours get to define gradients more)
plot(ff,col=fcol(ff,1:6,imp.weight = FALSE)) #useless X5 and X6 appear more colourful

#insert outlier in data set in X1 and X2
ff$X[1,1] = 10; ff$X[1,2] = 10

plot(ff,col=fcol(ff,1)) #colour not distorted, default: outlier.lim=3
plot(ff,col=fcol(ff,1,outlier.lim = Inf)) #colour gradient distorted by outlier 
plot(ff,col=fcol(ff,1,outlier.lim = 0.5)) #too little outlier.lim 


## End(Not run)

Warning messages:
1: In rgl.init(initValue, onlyNULL) : RGL: unable to open X11 display
2: 'rgl_init' failed, running with rgl.useNULL = TRUE 
3: .onUnload failed in unloadNamespace() for 'rgl', details:
  call: fun(...)
  error: object 'rgl_quit' not found 
randomForest 4.6-14
Type rfNews() to see new features/changes/bug fixes.
[1] "compute goodness-of-fit with leave-one-out k-nearest neighbor(guassian weighting), kknn package"
[1] "compute goodness-of-fit with leave-one-out k-nearest neighbor(guassian weighting), kknn package"
[1] "compute goodness-of-fit with leave-one-out k-nearest neighbor(guassian weighting), kknn package"
[1] "no $FCfit found, computing tempoary LOO-kNN-gaussion fit to main affect"
[1] "use ff = convolute_ff(ff) to compute a fixed fit"
[1] "compute goodness-of-fit with leave-one-out k-nearest neighbor(guassian weighting), kknn package"
[1] "compute goodness-of-fit with leave-one-out k-nearest neighbor(guassian weighting), kknn package"
[1] "compute goodness-of-fit with leave-one-out k-nearest neighbor(guassian weighting), kknn package"
[1] "compute goodness-of-fit with leave-one-out k-nearest neighbor(guassian weighting), kknn package"
[1] "compute goodness-of-fit with leave-one-out k-nearest neighbor(guassian weighting), kknn package"
[1] "compute goodness-of-fit with leave-one-out k-nearest neighbor(guassian weighting), kknn package"