This may be a little premature, but...
\nThe following relatively simple svm approach may be of interest to some people who, like me, are relatively new to Kaggle. A variant of it placed me 83rd but the private leaderboard suggests it could have been closer to top 20 if I hadn't 'tuned' it further in the final week.
\nThe CV and public leaderboard scores were both close to 0.40. I'll return with the CV code once I know which of my 15 submissions placed where (I'm assuming we learn that at some stage?).
\nMy thanks to BreakfastPirate, Abhishek and all for their valuable forum posts. I'll be looking out for further advice on how people selected their best models for the final entries.
\nR code:
\nsetwd("~/Documents/#RStudio/Africa Soil")
library(e1071); library(prospectr)
# Data preparation; first derivative of half of spectra via SG (all the 'good stuff' seems to be in that half)
myData <- read.csv("training.csv", header = TRUE)
sg21 <- savitzkyGolay(myData[ , c(1800:3579)], p = 3, w = 21, m = 1)
sg21 <- data.frame(sg21)
# Final models (after much CV exploration of parameters)
svm.Ca <- svm(sg21, myData$Ca, cost = 50, gam = 0.00005, ep = 0.01)
svm.SOC <- svm(sg21, myData$SOC, cost = 40, gam = 0.00005, ep = 0.05)
svm.pH <- svm(sg21, myData$pH, cost = 30, gam = 0.0001, ep = 0.15)
svm.P <- svm(sg21, myData$P, cost = 10, gam = 0.0001, ep = 0.1)
svm.Sand <- svm(sg21, myData$Sand, cost = 5, gam = 0.0001, ep = 0.12)
# Read test data and make predictions
testX <- read.csv("sorted_test.csv", header = TRUE)
test.sg21 <- data.frame(savitzkyGolay(testX[, c(1800:3579)],p = 3, w = 21, m = 1))
sub <- read.csv("sub Oct-04 sg21.csv", header = TRUE)
sub$Ca <- predict(svm.Ca, test.sg21)
sub$SOC <- predict(svm.SOC, test.sg21)
sub$pH <- predict(svm.pH, test.sg21)
sub$P <- predict(svm.P, test.sg21)
sub$Sand <- predict(svm.Sand, test.sg21)
write.csv(sub, file = "sub xxxxx.csv", row.names = FALSE)
This may be a little premature, but...
\nThe following relatively simple svm approach may be of interest to some people who, like me, are relatively new to Kaggle. A variant of it placed me 83rd but the private leaderboard suggests it could have been closer to top 20 if I hadn't 'tuned' it further in the final week.
\nThe CV and public leaderboard scores were both close to 0.40. I'll return with the CV code once I know which of my 15 submissions placed where (I'm assuming we learn that at some stage?).
\nMy thanks to BreakfastPirate, Abhishek and all for their valuable forum posts. I'll be looking out for further advice on how people selected their best models for the final entries.
\nR code:
\nsetwd("~/Documents/#RStudio/Africa Soil")
library(e1071); library(prospectr)
# Data preparation; first derivative of half of spectra via SG (all the 'good stuff' seems to be in that half)
myData <- read.csv("training.csv", header = TRUE)
sg21 <- savitzkyGolay(myData[ , c(1800:3579)], p = 3, w = 21, m = 1)
sg21 <- data.frame(sg21)
# Final models (after much CV exploration of parameters)
svm.Ca <- svm(sg21, myData$Ca, cost = 50, gam = 0.00005, ep = 0.01)
svm.SOC <- svm(sg21, myData$SOC, cost = 40, gam = 0.00005, ep = 0.05)
svm.pH <- svm(sg21, myData$pH, cost = 30, gam = 0.0001, ep = 0.15)
svm.P <- svm(sg21, myData$P, cost = 10, gam = 0.0001, ep = 0.1)
svm.Sand <- svm(sg21, myData$Sand, cost = 5, gam = 0.0001, ep = 0.12)
# Read test data and make predictions
testX <- read.csv("sorted_test.csv", header = TRUE)
test.sg21 <- data.frame(savitzkyGolay(testX[, c(1800:3579)],p = 3, w = 21, m = 1))
sub <- read.csv("sub Oct-04 sg21.csv", header = TRUE)
sub$Ca <- predict(svm.Ca, test.sg21)
sub$SOC <- predict(svm.SOC, test.sg21)
sub$pH <- predict(svm.pH, test.sg21)
sub$P <- predict(svm.P, test.sg21)
sub$Sand <- predict(svm.Sand, test.sg21)
write.csv(sub, file = "sub xxxxx.csv", row.names = FALSE)