Code and data for Super-organised som using kohonen package

ssomdata.csv

supersom.R

library(kohonen)
library(sqldf)

# Load data, available at https://gist.github.com/ottadini/6913696
somdata <- read.csv("ssomdata.csv")
# This data also includes a depth marker (which can be ignored), and classifications by way of Stratigraphy. Strat is not a defining label for the target variable MEAS_TC, but is a guide at least. 

# Create SCALED test and training sets from data:
unscale.cols <- c("Strat", "DEPTH")
scale.cols <- setdiff(colnames(somdata), c("Strat", "DEPTH"))

scaledsom <- somdata
scaledsom[, scale.cols] <- scale(somdata[, scale.cols])

inTrain <- sample(nrow(scaledsom), nrow(scaledsom)*(2/3))
training <- scaledsom[inTrain, ]
testing <- scaledsom[-inTrain, ]

# Following the examples in Wehrens and Buydens, 2007, 21(5), J Stat Soft.

# Super-organised kohonen map, requiring data in a list of data frames
# So ugly all this wrangling
somX <- list()
somX$training <- as.matrix(training[, -c(1, 2, 3)])  # exclude Strat, DEPTH, MEAS_TC
somX$Strat <- as.matrix(training[, "Strat"])
somX$MeasTC <- as.matrix(training[, "MEAS_TC"])

somXtest <- list()
somXtest$newdata <- as.matrix(testing[, -c(1, 2, 3)])
somXtest$Strat <- as.matrix(testing[, "Strat"])
somXtest$MeasTC <- as.matrix(testing[, "MEAS_TC"])

ssom <- supersom(data=somX, whatmap=1, contin=TRUE, rlen=500, 
                   grid=somgrid(5, 5, "hexagonal"))

# Figure 1. Mapping of all observations
plot(ssom, type="mapping", pch=1)

ssom.prediction <- predict(ssom, newdata=somXtest$newdata, 
                           trainY=somX$MeasTC)

predictions <- ssom.prediction$prediction
observed <- somXtest$MeasTC

fit <- lm(predictions ~ observed)
plot(predictions ~ observed)
abline(fit)