KaggleDays SF: H2O AutoML solution

kaggledays-sf_h2o_automl_6000.R

### Kaggle Days SF: Hackathon submission (8th place) 

# I used the latest version of H2O (3.24.0.1)
# Latest stable always here: http://h2o-release.s3.amazonaws.com/h2o/latest_stable.html
# H2O 3.24.0.1: http://h2o-release.s3.amazonaws.com/h2o/rel-yates/1/index.html
# If you are a Python user, you can use the demo Python code available on the H2O AutoML User Guide 
# instead: http://docs.h2o.ai/h2o/latest-stable/h2o-docs/automl.html

# Unfortunately it was a private competition, so the data is not publicly available! 

library(h2o)

h2o.init(max_mem_size = "40G")

# Import training data
train_file <- "/home/ledell/code/kaggle/kaggledays-sf-hackathon/input/train.csv"
train <- h2o.importFile(train_file)

# Identify predictors and response
y <- "target"
x <- setdiff(names(train), c(y, "id"))

# For binary classification, response should be a factor
train[,y] <- as.factor(train[,y])

# Train H2O AutoML for 6000 seconds (100 mins)
# If you want the same result that I had (43 models + 2 ensembles), 
# set max_models = 43, max_runtime_secs = 9999999 (unlimited).  
# Since you may have more/fewer cores than I do on my machine, 
# running for 100 mins on your hardware may produce more/fewer 
# models in the same amount of time.
aml <- h2o.automl(x = x, y = y, 
                  training_frame = train, 
                  seed = 1,
                  max_runtime_secs = 6000)

# When it finishes, check out the leaderboard
print(aml@leaderboard)

# Import test data
test_file <- "/home/ledell/code/kaggle/kaggledays-sf-hackathon/input/test.csv"
test <- h2o.importFile(test_file)
test_id <- test[,"id"]  #save for submission file
test <- test[, x]

# Predict on the test set
pred <- predict(aml, test)

# Create submission file
submission <- h2o.cbind(test_id, pred[,"p1"])
names(submission) <- c("id", "target")
sub_file <- "/home/ledell/code/kaggle/kaggledays-sf-hackathon/submissions/h2o_automl_6000.csv"
h2o.exportFile(submission, path = sub_file)


# Here is a snapshot (head) of the leaderboard:

# > aml@leaderboard
# model_id       auc    logloss
# 1 StackedEnsemble_BestOfFamily_AutoML_20190411_152415 0.6476812 0.03137691
# 2    StackedEnsemble_AllModels_AutoML_20190411_152415 0.6447939 0.03134611
# 3       XGBoost_grid_1_AutoML_20190411_152415_model_5 0.6439845 0.03133052
# 4                    XGBoost_1_AutoML_20190411_152415 0.6430735 0.03134268
# 5          GBM_grid_1_AutoML_20190411_152415_model_10 0.6428834 0.03140852
# 6       XGBoost_grid_1_AutoML_20190411_152415_model_1 0.6413726 0.03128838
# mean_per_class_error       rmse         mse
# 1            0.4699024 0.07126896 0.005079264
# 2            0.4801189 0.07128310 0.005081280
# 3            0.4835419 0.07120733 0.005070484
# 4            0.4780277 0.07121024 0.005070898
# 5            0.4859112 0.07120488 0.005070135
# 6            0.4867847 0.07120804 0.005070586
#
# [45 rows x 6 columns]