shipunov (version 1.13)

Gradd: Classification grid and decision boundaries

Description

Adds to the 2D ordination either colored points to make classification grid, or lines to show decision boundaries

Usage

Gradd(model2var, data2var, spacing=75, what="points",
 trnsp=0.2, pch=16, cex=0.8, lwd=2, lty=2, lcol="grey", palette=NULL,
 type="ids", User.Predict=function(model2var, X) {}, ...)

Arguments

model2var

Model based on 'data2var' (see below).

data2var

Data with _exactly_ 2 variables, e.g., result of PCA.

spacing

Density of points to predict.

what

What to draw: either "points" for classification grid, or "lines" for decision boundaries

trnsp

Transparency of points.

pch

Type of points.

cex

Scale of points.

lwd

Width of lines.

lty

Type of lines.

lcol

Color of lines.

palette

Palette to use.

type

Type of the model: "ids", "lda", "tree", or "user" (see examples).

User.Predict

Function to define in case of 'type="user"'.

...

Additional arguments to points() or contour().

Details

Gradd() takes model and its 2D data, makes new data with the same range but made of dense equidistantly spaced (grid-like) points, then predicts class labels from this new data, probably also calculates decision boundaries, and finally plots either points colored by predition, or lines along boundaries.

Before you run Gradd(), make the model. This model should have 'predict' method, and use ids (to make colors) and exactly 2 variables with names same as 'data2var' column names, e.g:

model2var <- somemodel(ids ~ ., data=cbind(ids, data2var))

If the model type is "user", the Gradd() uses predefined 'User.Predict(model2var, X)' function which must return factor ids from testing X data (see examples).

To plot both lines and grid, use Gradd() twice.

Gradd() is mainly a teching demo. It is useful if the goal is to illustrate the general properties of the supervised method and/or underlying data. It uses the entire 2D dataset to learn new data but learning from training subset is also possible, see the Naive Bayes example below.

Examples

# NOT RUN {
## SVM:
library(e1071)
iris.p <- prcomp(iris[, -5])$x[, 1:2]
iris.svm.pca <- svm(Species ~ ., data=cbind(iris[5], iris.p))
plot(iris.p, type="n", main="SVM")
Gradd(iris.svm.pca, iris.p) # type="ids" (default)
text(iris.p, col=as.numeric(iris[, 5]), labels=abbreviate(iris[, 5], 1,
 method="both.sides"))

## LDA:
library(MASS)
iris.p <- prcomp(iris[, -5])$x[, 1:2]
iris.lda.pca <- lda(Species ~ . , data=cbind(iris[5], iris.p))
plot(iris.p, type="n", main="LDA")
Gradd(iris.lda.pca, iris.p, type="lda")
text(iris.p, col=as.numeric(iris[, 5]), labels=abbreviate(iris[, 5], 1,
 method="both.sides"))

## tree::tree() (note how to draw decision boundaries):
library(tree)
iris.p <- prcomp(iris[, -5])$x[, 1:2]
iris.tree.pca <- tree(Species ~ . , data=cbind(iris[5], iris.p))
plot(iris.p, type="n", main="tree")
Gradd(iris.tree.pca, iris.p, type="tree", what="lines")
text(iris.p, col=as.numeric(iris[, 5]), labels=abbreviate(iris[, 5], 1,
 method="both.sides"))

## randomForest:
library(randomForest)
iris.p <- prcomp(iris[, -5])$x[, 1:2]
iris.rf.pca <- randomForest(Species ~ ., data=cbind(iris[5], iris.p))
plot(iris.p, type="n", main="randomForest")
Gradd(iris.rf.pca, iris.p) # type="ids" (default)
text(iris.p, col=as.numeric(iris[, 5]), labels=abbreviate(iris[, 5], 1,
 method="both.sides"))

## naiveBayes (note how to use training subsample):
library(e1071)
iris.p <- prcomp(iris[, -5])$x[, 1:2]
sel <- 1:nrow(iris) 
# }
# NOT RUN {
<!-- %in% seq(1, nrow(iris), 5) -->
# }
# NOT RUN {
plot(iris.p, col=iris$Species, pch=ifelse(sel, 19, 1), main="naiveBayes")
iris.nb2 <- naiveBayes(Species ~ ., data=cbind(iris[5], iris.p)[sel, ])
Gradd(iris.nb2, iris.p[sel, ], what="lines")

## rpart (note how to use MDS for the base plot):
iris.dist <- dist(iris[, -5], method="manhattan")
iris.dist[iris.dist == 0] <- abs(jitter(0))
library(MASS)
iris.m <- isoMDS(iris.dist)$points
colnames(iris.m) <- c("Dim1", "Dim2")
library(rpart)
iris.rpart.mds <- rpart(Species ~ . , data=cbind(iris[5], iris.m))
plot(iris.m, type="n", main="rpart + MDS")
Gradd(iris.rpart.mds, iris.m, type="tree")
text(iris.m, col=as.numeric(iris[, 5]), labels=abbreviate(iris[, 5], 1,
 method="both.sides"))

## QDA:
library(MASS)
iris.p <- prcomp(iris[, -5])$x[, 1:2]
iris.qda.pca <- qda(Species ~ . , data=cbind(iris[5], iris.p))
plot(iris.p, type="n", main="QDA")
Gradd(iris.qda.pca, iris.p, type="lda")
text(iris.p, col=as.numeric(iris[, 5]), labels=abbreviate(iris[, 5], 1,
 method="both.sides"))

## nnet:
library(nnet)
iris.p <- prcomp(iris[, -5])$x[, 1:2]
iris.nnet.pca <- nnet(Species ~ . , data=cbind(iris[5], iris.p), size=4)
plot(iris.p, type="n", main="nnet")
Gradd(iris.nnet.pca, iris.p, type="tree")
text(iris.p, col=as.numeric(iris[, 5]), labels=abbreviate(iris[, 5], 1,
 method="both.sides"))

## kNN (note how to employ User.Predict()):
library(class)
iris.p <- prcomp(iris[, -5])$x[, 1:2]
plot(iris.p, type="n", main="kNN")
Gradd(cbind(iris[5], iris.p), iris.p, type="user",
 User.Predict=function(model2var, X) knn(model2var[, 2:3], X, model2var[, 1], k=5))
text(iris.p, col=as.numeric(iris[, 5]), labels=abbreviate(iris[, 5], 1,
 method="both.sides"))
# }