init (GLRM, K-Means)

  • Available in: GLRM, K-means

  • Hyperparameter: yes

Description

This option specifies the initialization mode used in K-Means and GLRM. The options are Random, Furthest, PlusPlus, and User.

  • Random: Choose \(K\) clusters from the set of \(N\) observations at random so that each observation has an equal chance of being chosen.

  • Furthest (Default):

    1. Choose one center \(m_{1}\) at random.

    2. Calculate the difference between \(m_{1}\) and each of the remaining \(N-1\) observations \(x_{i}\). \(d(x_{i}, m_{1}) = ||(x_{i}-m_{1})||^2\)

    3. Choose \(m_{2}\) to be the \(x_{i}\) that maximizes \(d(x_{i}, m_{1})\).

    4. Repeat until \(K\) centers have been chosen.

  • PlusPlus:

    1. Choose one center \(m_{1}\) at random.

    2. Calculate the difference between \(m_{1}\) and each of the remaining \(N-1\) observations \(x_{i}\). \(d(x_{i}, m_{1}) = \|(x_{i}-m_{1})\|^2\)

    3. Let \(P(i)\) be the probability of choosing \(x_{i}\) as \(m_{2}\). Weight \(P(i)\) by \(d(x_{i}, m_{1})\) so that those \(x_{i}\) furthest from \(m_{2}\) have a higher probability of being selected than those \(x_{i}\) close to \(m_{1}\).

    4. Choose the next center \(m_{2}\) by drawing at random according to the weighted probability distribution.

    5. Repeat until \(K\) centers have been chosen.

  • User initialization allows you to specify a file (using the user_points parameter) that includes a vector of initial cluster centers.

Notes:

  • The user-specified points dataset must have the same number of columns as the training observations.

  • This option is ignored when estimate_k is enabled. In this case, the algorithm is deterministic.

  • If this option is not specified but a user-points file is specified, then this value will default to user.

Example

library(h2o)
h2o.init()

# import the seeds dataset:
# this dataset looks at three different types of wheat varieties
# the original dataset can be found at http://archive.ics.uci.edu/ml/datasets/seeds
seeds <- h2o.importFile("https://s3.amazonaws.com/h2o-public-test-data/smalldata/flow_examples/seeds_dataset.txt")

# set the predictor names
# ignore the 8th column which has the prior known clusters for this dataset
predictors <- colnames(seeds)[-length(seeds)]

# split into train and validation
seeds_splits <- h2o.splitFrame(data = seeds, ratios = 0.8, seed = 1234)
train <- seeds_splits[[1]]
valid <- seeds_splits[[2]]

# try using the `init` parameter:
# build the model with three clusters
seeds_kmeans <- h2o.kmeans(x = predictors, k = 3, init='Furthest', training_frame = train, validation_frame = valid, seed = 1234)

# print the total within cluster sum-of-square error for the validation dataset
print(paste0("Total sum-of-square error for valid dataset: ", h2o.tot_withinss(object = seeds_kmeans, valid = T)))


# select the values for `init` to grid over:
# Note: this dataset is too small to see significant differences between these options
# the purpose of the example is to show how to use grid search with `init` if desired
hyper_params <- list( init = c("PlusPlus", "Furthest", "Random")  )

# this example uses cartesian grid search because the search space is small
# and we want to see the performance of all models. For a larger search space use
# random grid search instead: list(strategy = "RandomDiscrete")
grid <- h2o.grid(x = predictors, k = 3, training_frame = train, validation_frame = valid,
                 algorithm = "kmeans", grid_id = "seeds_grid", hyper_params = hyper_params,
                 search_criteria = list(strategy = "Cartesian"), seed = 1234)

## Sort the grid models by TotSS
sorted_grid <- h2o.getGrid("seeds_grid", sort_by  = "tot_withinss", decreasing = F)
sorted_grid
import h2o
from h2o.estimators.kmeans import H2OKMeansEstimator
h2o.init()

# import the seeds dataset:
# this dataset looks at three different types of wheat varieties
# the original dataset can be found at http://archive.ics.uci.edu/ml/datasets/seeds
seeds = h2o.import_file("https://s3.amazonaws.com/h2o-public-test-data/smalldata/flow_examples/seeds_dataset.txt")

# set the predictor names
# ignore the 8th column which has the prior known clusters for this dataset
predictors = seeds.columns[0:7]

# split into train and validation sets
train, valid = seeds.split_frame(ratios = [.8], seed = 1234)

# try using the `init` parameter:
# initialize the estimator then train the model
seeds_kmeans = H2OKMeansEstimator(k = 3, init='Furthest', seed = 1234)
seeds_kmeans.train(x = predictors, training_frame = train, validation_frame= valid)

# print the total within cluster sum-of-square error for the validation dataset
print("sum-of-square error for valid:",seeds_kmeans.tot_withinss(valid = True))


# grid over `init`
# import Grid Search
from h2o.grid.grid_search import H2OGridSearch

# select the values for `init` to grid over
# Note: this dataset is too small to see significant differences between these options
# the purpose of the example is to show how to use grid search with `init` if desired
hyper_params = {'init': ["PlusPlus", "Furthest", "Random"]}

# this example uses cartesian grid search because the search space is small
# and we want to see the performance of all models. For a larger search space use
# random grid search instead: {'strategy': "RandomDiscrete"}
# initialize the estimator
seeds_kmeans = H2OKMeansEstimator(k = 3, seed = 1234)

# build grid search with previously made Kmeans and hyperparameters
grid = H2OGridSearch(model = seeds_kmeans, hyper_params = hyper_params,
                     search_criteria = {'strategy': "Cartesian"})

# train using the grid
grid.train(x = predictors, training_frame = train, validation_frame = valid)

# sort the grid models by total within cluster sum-of-square error.
sorted_grid = grid.get_grid(sort_by='tot_withinss', decreasing= False)
print(sorted_grid)