H2O K-Means Auto-estimate K (wine data demo)

h2o_kmeans_estimate_k_wine.R

# H2O's K-Means algo can estimate the optimal number of clusters (method by Leland Wilkinson)
# http://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/k-means.html#estimating-k-in-k-means
#
# This demo is an extension of Kasia's blog post here: 
# https://kkulma.github.io/2017-04-24-determining-optimal-number-of-clusters-in-your-data/

library(rattle)  # wine data

# Remove the factor col & convert to an H2O Frame
# Note: You can skip the scale() here since H2O K-Means standardizes automatically
data(wine)
wine <- as.h2o(wine[,-1]) 

h2o.init() # Start a local H2O Cluster

# Train a H2O K-Means model, auto-estimate best value for k  (note: k here is maximum possible k)
fit <- h2o.kmeans(training_frame = wine, k = 20, estimate_k = TRUE)

# It found 3 clusters to be optimal
print(fit)
# Model Details:
# ==============
#   
# H2OClusteringModel: kmeans
# Model ID:  KMeans_model_R_1503103871072_10 
# Model Summary: 
#   number_of_rows number_of_clusters number_of_categorical_columns
# 1            178                  3                             0
# number_of_iterations within_cluster_sum_of_squares
# 1                   19                    1270.74912
# total_sum_of_squares between_cluster_sum_of_squares
# 1           2301.00000                     1030.25088
# 
# 
# H2OClusteringMetrics: kmeans
# ** Reported on training data. **
#   
#   
#   Total Within SS:  1270.749
# Between SS:  1030.251
# Total SS:  2301 
# Centroid Statistics: 
#   centroid     size within_cluster_sum_of_squares
# 1        1 51.00000                     326.35370
# 2        2 65.00000                     558.69710
# 3        3 62.00000                     385.69830