outliers impact on neural net classifier

nnet_outliers.R

library(nnet)
library(h2o)
h2o.init()

set.seed(123)

n <- 1000

x1 <- runif(n)
x2 <- runif(n)

z <- 30*(x2 - x1)        
p <-  1/(1+exp(-z)) 
y <- rbinom(n,1,p)

x1t <- seq(0,1,1/20)
x2t <- seq(0,1,1/20)
dt <- expand.grid(x1=x1t,x2=x2t)
dxt <- as.h2o(dt)

par(pty="s")
plot(x2 ~ x1, col = ifelse(y==0,"blue","red"))

d <- data.frame(x1,x2,y)
## outliers ON/OFF:
d <- rbind(d, data.frame(x1=1000,x2=0,y=1))
dx <- as.h2o(d)

## choose model glm/nnet/h2o.DL:
#md <- glm(y~., d, family=binomial)
#md <- nnet(y~., d, size = 10, decay = 0.1, maxit = 1000)
md <- h2o.deeplearning(x=1:2, y=3, training_frame=dx,
       hidden = 10, activation = "Rectifier", epochs = 10)
## try with tanh, try with CV (auto early stopping)

## choose d/dx for R/h2o
#yt <- matrix(predict(md, dt),21,21)
yt <- matrix(h2o.predict(md, dxt),21,21)
contour(x=x1t, y=x2t, z=yt, levels=0.5, add=TRUE, lwd=3)