Training Models¶
H2O supports training of supervised models (where the outcome variable is known) and unsupervised models (unlabeled data). Below we present examples of classification, regression, clustering, dimensionality reduction and training on data segments (train a set of models – one for each partition of the data).
Supervised Learning¶
Supervised learning algorithms support classification and regression problems.
Classification and Regression¶
In classification problems, the output or “response” variable is a categorical value. The answer can be binary (for example, yes/no), or it can be multiclass (for example, dog/cat/mouse).
In regression problems, the output or “response” variable is a numeric value. An example would be predicting someone’s age or the sale price of a house.
Classification vs Regression in H2O models: The way that you tell H2O whether you want to do classification versus regression for a particular supervised algorithm is by encoding the response column as either a categorical/factor type (classification) or a numeric type (regression). If your column is represented as strings (“yes”, “no”), then H2O will automatically encode that column as an categorical/factor (aka. “enum”) type when you import your dataset. However if you have a column with integers that represents a class in a classification problem (0, 1), you will have to change the column type from numeric to categorical/factor (aka. “enum”) type. The reason that H2O requires the response column to be encoded as the “correct” type for a particular task is to maximze efficiency of the algorithm.
Classification Example¶
This example uses the Prostate dataset and H2O’s GLM algorithm to predict the likelihood of a patient being diagnosed with prostate cancer. The dataset includes the following columns:
ID: A row identifier. This can be dropped from the list of predictors.
CAPSULE: Whether the tumor penetrated the prostatic capsule
AGE: The patient’s age
RACE: The patient’s race
DPROS: The result of the digital rectal exam, where 1=no nodule; 2=unilober nodule on the left; 3 =unilibar nodule on the right; and 4=bilobar nodule.
DCAPS: Whether there existed capsular involvement on the rectal exam
PSA: The Prostate Specific Antigen Value (mg/ml)
VOL: The tumor volume (cm3)
GLEASON: The patient’s Gleason score in the range 0 to 10
This example uses only the AGE, RACE, VOL, and GLEASON columns to make the prediction.
library(h2o)
h2o.init()
# import the prostate dataset
df <- h2o.importFile("https://h2o-public-test-data.s3.amazonaws.com/smalldata/prostate/prostate.csv")
# convert columns to factors
df$CAPSULE <- as.factor(df$CAPSULE)
df$RACE <- as.factor(df$RACE)
df$DCAPS <- as.factor(df$DCAPS)
df$DPROS <- as.factor(df$DPROS)
# set the predictor and response columns
predictors <- c("AGE", "RACE", "VOL", "GLEASON")
response <- "CAPSULE"
# split the dataset into train and test sets
df_splits <- h2o.splitFrame(data = df, ratios = 0.8, seed = 1234)
train <- df_splits[[1]]
test <- df_splits[[2]]
# build a GLM model
prostate_glm <- h2o.glm(family = "binomial",
x = predictors,
y = response,
training_frame = df,
lambda = 0,
compute_p_values = TRUE)
# predict using the GLM model and the testing dataset
predict <- h2o.predict(object = prostate_glm, newdata = test)
# view a summary of the predictions
h2o.head(predict)
predict p0 p1 StdErr
1 1 0.5318993 0.46810066 0.4472349
2 0 0.7269800 0.27302003 0.4413739
3 0 0.9009476 0.09905238 0.3528526
4 0 0.9062937 0.09370628 0.2874410
5 1 0.6414760 0.35852398 0.2719742
6 1 0.5092692 0.49073080 0.4007240
import h2o
h2o.init()
from h2o.estimators.glm import H2OGeneralizedLinearEstimator
# import the prostate dataset
prostate = h2o.import_file("https://h2o-public-test-data.s3.amazonaws.com/smalldata/prostate/prostate.csv")
# convert columns to factors
prostate['CAPSULE'] = prostate['CAPSULE'].asfactor()
prostate['RACE'] = prostate['RACE'].asfactor()
prostate['DCAPS'] = prostate['DCAPS'].asfactor()
prostate['DPROS'] = prostate['DPROS'].asfactor()
# set the predictor and response columns
predictors = ["AGE", "RACE", "VOL", "GLEASON"]
response_col = "CAPSULE"
# split into train and testing sets
train, test = prostate.split_frame(ratios = [0.8], seed = 1234)
# set GLM modeling parameters
# and initialize model training
glm_model = H2OGeneralizedLinearEstimator(family= "binomial",
lambda_ = 0,
compute_p_values = True)
glm_model.train(predictors, response_col, training_frame= prostate)
# predict using the model and the testing dataset
predict = glm_model.predict(test)
# View a summary of the prediction
predict.head()
predict p0 p1 StdErr
--------- -------- --------- --------
1 0.531899 0.468101 0.447235
0 0.72698 0.27302 0.441374
0 0.900948 0.0990524 0.352853
0 0.906294 0.0937063 0.287441
1 0.641476 0.358524 0.271974
1 0.509269 0.490731 0.400724
1 0.355024 0.644976 0.235607
1 0.304671 0.695329 1.33002
1 0.472833 0.527167 0.170934
0 0.720066 0.279934 0.221276
[10 rows x 4 columns]
Regression Example¶
This example uses the Boston Housing data and H2O’s GLM algorithm to predict the median home price using all available features. The dataset includes the following columns:
crim: The per capita crime rate by town
zn: The proportion of residential land zoned for lots over 25,000 sq.ft
indus: The proportion of non-retail business acres per town
chas: A Charles River dummy variable (1 if the tract bounds the Charles river; 0 otherwise)
nox: Nitric oxides concentration (parts per 10 million)
rm: The average number of rooms per dwelling
age: The proportion of owner-occupied units built prior to 1940
dis: The weighted distances to five Boston employment centers
rad: The index of accessibility to radial highways
tax: The full-value property-tax rate per $10,000
ptratio: The pupil-teacher ratio by town
b: 1000(Bk - 0.63)^2, where Bk is the black proportion of population
lstat: The % lower status of the population
medv: The median value of owner-occupied homes in $1000’s
library(h2o)
h2o.init()
# import the boston dataset:
# this dataset looks at features of the boston suburbs and predicts median housing prices
# the original dataset can be found at https://archive.ics.uci.edu/ml/datasets/Housing
boston <- h2o.importFile("https://s3.amazonaws.com/h2o-public-test-data/smalldata/gbm_test/BostonHousing.csv")
# set the predictor names and the response column name
predictors <- colnames(boston)[1:13]
# this example will predict the medv column
# you can run the following to see that medv is indeed a numeric value
h2o.isnumeric(boston["medv"])
[1] TRUE
# set the response column to "medv", which is the median value of owner-occupied homes in $1000's
response <- "medv"
# convert the `chas` column to a factor
# `chas` = Charles River dummy variable (= 1 if tract bounds river; 0 otherwise)
boston["chas"] <- as.factor(boston["chas"])
# split into train and test sets
boston_splits <- h2o.splitFrame(data = boston, ratios = 0.8, seed = 1234)
train <- boston_splits[[1]]
test <- boston_splits[[2]]
# set the `alpha` parameter to 0.25 and train the model
boston_glm <- h2o.glm(x = predictors,
y = response,
training_frame = train,
alpha = 0.25)
# predict using the GLM model and the testing dataset
predict <- h2o.predict(object = boston_glm, newdata = test)
# view a summary of the predictions
h2o.head(predict)
predict
1 28.29427
2 19.45689
3 19.08230
4 16.90933
5 16.23141
6 18.23614
import h2o
from h2o.estimators.glm import H2OGeneralizedLinearEstimator
h2o.init()
# import the boston dataset:
# this dataset looks at features of the boston suburbs and predicts median housing prices
# the original dataset can be found at https://archive.ics.uci.edu/ml/datasets/Housing
boston = h2o.import_file("https://s3.amazonaws.com/h2o-public-test-data/smalldata/gbm_test/BostonHousing.csv")
# set the predictor columns
predictors = boston.columns[:-1]
# this example will predict the medv column
# you can run the following to see that medv is indeed a numeric value
boston["medv"].isnumeric()
[True]
# set the response column to "medv", which is the median value of owner-occupied homes in $1000's
response = "medv"
# convert the `chas` column to a factor
# `chas` = Charles River dummy variable (= 1 if tract bounds river; 0 otherwise)
boston['chas'] = boston['chas'].asfactor()
# split into train and testing sets
train, test = boston.split_frame(ratios = [0.8], seed = 1234)
# set the `alpha` parameter to 0.25
# then initialize the estimator then train the model
boston_glm = H2OGeneralizedLinearEstimator(alpha = 0.25)
boston_glm.train(x = predictors,
y = response,
training_frame = train)
# predict using the model and the testing dataset
predict = boston_glm.predict(test)
# View a summary of the prediction
predict.head()
predict
---------
28.2943
19.4569
19.0823
16.9093
16.2314
18.2361
12.6945
17.5583
15.4797
20.7294
[10 rows x 1 column]
Unsupervised Learning¶
Unsupervised learning algorithms include clustering and anomaly detection methods. Unsupervised learning algorithms such as GLRM and PCA can also be used to perform dimensionality reduction.
Clustering Example¶
The example below uses the K-Means algorithm to build a simple clustering model of the Iris dataset.
library(h2o)
h2o.init()
# Import the iris dataset into H2O:
iris <- h2o.importFile("http://h2o-public-test-data.s3.amazonaws.com/smalldata/iris/iris_wheader.csv")
# Set the predictors:
predictors <- c("sepal_len", "sepal_wid", "petal_len", "petal_wid")
# Split the dataset into a train and valid set:
iris_split <- h2o.splitFrame(data = iris, ratios = 0.8, seed = 1234)
train <- iris_split[[1]]
valid <- iris_split[[2]]
# Build and train the model:
iris_kmeans <- h2o.kmeans(k = 10,
estimate_k = TRUE,
standardize = FALSE,
seed = 1234,
x = predictors,
training_frame = train,
validation_frame = valid)
# Eval performance:
perf <- h2o.performance(iris_kmeans)
perf
H2OClusteringMetrics: kmeans
** Reported on training data. **
Total Within SS: 63.09516
Between SS: 483.8141
Total SS: 546.9092
Centroid Statistics:
centroid size within_cluster_sum_of_squares
1 1 36.00000 11.08750
2 2 51.00000 30.78627
3 3 36.00000 21.22139
import h2o
from h2o.estimators import H2OKMeansEstimator
h2o.init()
# Import the iris dataset into H2O:
iris = h2o.import_file("http://h2o-public-test-data.s3.amazonaws.com/smalldata/iris/iris_wheader.csv")
# Set the predictors:
predictors = ["sepal_len", "sepal_wid", "petal_len", "petal_wid"]
# Split the dataset into a train and valid set:
train, valid = iris.split_frame(ratios=[.8], seed=1234)
# Build and train the model:
iris_kmeans = H2OKMeansEstimator(k=10,
estimate_k=True,
standardize=False,
seed=1234)
iris_kmeans.train(x=predictors,
training_frame=train,
validation_frame=valid)
# Eval performance:
perf = iris_kmeans.model_performance()
perf
ModelMetricsClustering: kmeans
** Reported on train data. **
MSE: NaN
RMSE: NaN
Total Within Cluster Sum of Square Error: 63.09516069071749
Total Sum of Square Error to Grand Mean: 546.9092331233204
Between Cluster Sum of Square Error: 483.8140724326029
Centroid Statistics:
centroid size within_cluster_sum_of_squares
-- ---------- ------ -------------------------------
1 36 11.0875
2 51 30.7863
3 36 21.2214
Anomaly Detection Example¶
This example uses the Isolation Forest algorithm to detect anomalies in the Electrocardiograms (ECG) dataset.
library(h2o)
h2o.init()
# import the ecg discord datasets:
train <- h2o.importFile("http://s3.amazonaws.com/h2o-public-test-data/smalldata/anomaly/ecg_discord_train.csv")
test <- h2o.importFile("http://s3.amazonaws.com/h2o-public-test-data/smalldata/anomaly/ecg_discord_test.csv")
# train using the `sample_size` parameter:
isofor_model <- h2o.isolationForest(training_frame = train,
sample_size = 5,
ntrees = 7,
seed = 12345)
# test the predictions and retrieve the mean_length.
# mean_length is the average number of splits it took to isolate
# the record across all the decision trees in the forest. Records
# with a smaller mean_length are more likely to be anomalous
# because it takes fewer partitions of the data to isolate them.
pred <- h2o.predict(isofor_model, test)
pred
predict mean_length
1 0.5555556 1.857143
2 0.5555556 1.857143
3 0.3333333 2.142857
4 1.0000000 1.285714
5 0.7777778 1.571429
6 0.6666667 1.714286
[23 rows x 2 columns]
import h2o
from h2o.estimators.isolation_forest import H2OIsolationForestEstimator
h2o.init()
# import the ecg discord datasets:
train = h2o.import_file("http://s3.amazonaws.com/h2o-public-test-data/smalldata/anomaly/ecg_discord_train.csv")
test = h2o.import_file("http://s3.amazonaws.com/h2o-public-test-data/smalldata/anomaly/ecg_discord_test.csv")
# build a model using the `sample_size` parameter:
isofor_model = H2OIsolationForestEstimator(sample_size = 5, ntrees = 7, seed = 12345)
isofor_model.train(training_frame = train)
# test the predictions and retrieve the mean_length.
# mean_length is the average number of splits it took to isolate
# the record across all the decision trees in the forest. Records
# with a smaller mean_length are more likely to be anomalous
# because it takes fewer partitions of the data to isolate them.
pred = isofor_model.predict(test)
pred
predict mean_length
--------- -------------
0.555556 1.85714
0.555556 1.85714
0.333333 2.14286
1 1.28571
0.777778 1.57143
0.666667 1.71429
0.666667 1.71429
0 2.57143
0.888889 1.42857
0.777778 1.57143
[23 rows x 2 columns]
Dimensionality Reduction Example¶
Below is a simple example showing how to use the GLRM algorithm for dimensionality reduction of the USArrests dataset.
library(h2o)
h2o.init()
# Import the USArrests dataset into H2O:
arrests <- h2o.importFile("https://s3.amazonaws.com/h2o-public-test-data/smalldata/pca_test/USArrests.csv")
# Split the dataset into a train and valid set:
arrests_splits <- h2o.splitFrame(data = arrests, ratios = 0.8, seed = 1234)
train <- arrests_splits[[1]]
valid <- arrests_splits[[2]]
# Build and train the model:
glrm_model = h2o.glrm(training_frame = train,
k = 4,
loss = "Quadratic",
gamma_x = 0.5,
gamma_y = 0.5,
max_iterations = 700,
recover_svd = TRUE,
init = "SVD",
transform = "STANDARDIZE")
# Eval performance:
arrests_perf <- h2o.performance(glrm_model)
arrests_perf
H2ODimReductionMetrics: glrm
** Reported on training data. **
Sum of Squared Error (Numeric): 1.983347e-13
Misclassification Error (Categorical): 0
Number of Numeric Entries: 144
Number of Categorical Entries: 0
# Generate predictions on a validation set:
arrests_pred <- h2o.predict(glrm_model, newdata = valid)
arrests_pred
reconstr_Murder reconstr_Assault reconstr_UrbanPop reconstr_Rape
1 0.2710690 0.2568493 -1.08479880 -0.281431002
2 2.3535244 0.5080499 -0.39237403 0.357493436
3 -1.3270945 -1.3460498 -0.60010146 -1.113046938
4 1.8692325 0.9626034 0.02308083 -0.007606243
5 -1.3513091 -1.0230776 -1.01555632 -1.468004959
6 0.8764339 1.5726620 0.09232330 0.560326590
[14 rows x 4 columns]
import h2o
from h2o.estimators import H2OGeneralizedLowRankEstimator
h2o.init()
# Import the USArrests dataset into H2O:
arrestsH2O = h2o.import_file("https://s3.amazonaws.com/h2o-public-test-data/smalldata/pca_test/USArrests.csv")
# Split the dataset into a train and valid set:
train, valid = arrestsH2O.split_frame(ratios = [.8], seed = 1234)
# Build and train the model:
glrm_model = H2OGeneralizedLowRankEstimator(k = 4,
loss = "quadratic",
gamma_x = 0.5,
gamma_y = 0.5,
max_iterations = 700,
recover_svd = True,
init = "SVD",
transform = "standardize")
glrm_model.train(training_frame=train)
# Eval performance
arrests_perf = glrm_model.model_performance()
arrests_perf
ModelMetricsGLRM: glrm
** Reported on train data. **
MSE: NaN
RMSE: NaN
Sum of Squared Error (Numeric): 1.983347263428422e-13
Misclassification Error (Categorical): 0.0
# Generate predictions on a validation set:
pred = glrm_model.predict(valid)
pred
reconstr_Murder reconstr_Assault reconstr_UrbanPop reconstr_Rape
----------------- ------------------ ------------------- ---------------
0.271069 0.256849 -1.0848 -0.281431
2.35352 0.50805 -0.392374 0.357493
-1.32709 -1.34605 -0.600101 -1.11305
1.86923 0.962603 0.0230808 -0.00760624
-1.35131 -1.02308 -1.01556 -1.468
0.876434 1.57266 0.0923233 0.560327
-0.794373 -0.23359 1.33869 -0.605964
1.07015 1.03437 0.577021 1.30067
-0.818588 -0.795801 -0.253889 -0.585681
-0.0679354 -0.113971 1.61566 -0.352423
[14 rows x 4 columns]
Training Segments¶
In H2O, you can perform bulk training on segments, or partitions, of the training set. The train_segments() method in Python and h2o.train_segments() function in R train an H2O model for each segment (subpopulation) of the training dataset.
Defining a Segmented Model¶
The train_segments() function accepts the following parameters:
x: A list of column names or indices indicating the predictor columns.
y: An index or a column name indicating the response column.
algorithm (R only): When building a segmented model in R, specify the algorithm to use. Available options include:
aggregatorcoxphdeeplearninggamgbmglrmglmisolationforestkmeansnaivebayespcapsvmrandomForeststackedensemblesvdtargetencoderword2vecxgboost
training_frame: The H2OFrame having the columns indicated by
xandy(as well as any additional columns specified byfold_column,offset_column, andweights_column).offset_column: The name or index of the column in the
training_framethat holds the offsets.fold_column: The name or index of the column in the
training_framethat holds the per-row fold assignments.weights_column: The name or index of the column in the
training_framethat holds the per-row weights.validation_frame: The H2OFrame with validation data to be scored on while training.
max_runtime_secs: Maximum allowed runtime in seconds for each model training. Use 0 to disable. Please note that regardless of how this parameter is set, a model will be built for each input segment. This parameter only affects individual model training.
segments (Python)/segment_columns (R): A list of columns to segment by. H2O will group the training (and validation) dataset by the segment-by columns and train a separate model for each segment (group of rows). As an alternative to providing a list of columns, users can also supply an explicit enumeration of segments to build the models for. This enumeration needs to be represented as H2OFrame.
segment_models_id: Identifier for the returned collection of Segment Models. If not specified it will be automatically generated.
parallelism: Level of parallelism of the bulk segment models building. This is the maximum number of models each H2O node will be building in parallel.
verbose: Enable to print additional information during model building. Defaults to False.
Segmented Model Example¶
Below is a simple example describing how to train a segmented model. A more detailed example is available in the H2O Segment Model Building demo.
library(h2o)
h2o.init()
# import the titanic dataset
titanic <- h2o.importFile("https://s3.amazonaws.com/h2o-public-test-data/smalldata/gbm_test/titanic.csv")
# set the predictor and response columns
predictors <- c("name", "sex", "age", "sibsp", "parch", "ticket", "fare", "cabin")
response <- "survived"
# convert the response columnn to a factor
titanic['survived'] <- as.factor(titanic['survived'])
# split the dataset into training and validation datasets
titanic.splits <- h2o.splitFrame(data = titanic, ratios = .8, seed = 1234)
train <- titanic.splits[[1]]
valid <- titanic.splits[[2]]
# train a segmented model by iterating over the plcass column
titanic_models <- h2o.train_segments(algorithm = "gbm",
segment_columns = "pclass",
x = predictors,
y = response,
training_frame = train,
validation_frame = valid,
seed = 1234)
# convert the segmented models to an H2OFrame
as.data.frame(titanic_models)
import h2o
from h2o.estimators.gbm import H2OGradientBoostingEstimator
h2o.init()
# import the titanic dataset:
titanic = h2o.import_file("https://s3.amazonaws.com/h2o-public-test-data/smalldata/gbm_test/titanic.csv")
# set the predictor and response columns
predictors = ["name", "sex", "age", "sibsp", "parch", "ticket", "fare", "cabin"]
response = "survived"
# convert the response columnn to a factor
titanic[response] = titanic[response].asfactor()
# split the dataset into training and validation datasets
train, valid = titanic.split_frame(ratios = [.8], seed = 1234)
# train a segmented model by iterating over the plcass column
titanic_gbm = H2OGradientBoostingEstimator(seed = 1234)
titanic_models = titanic_gbm.train_segments(segments = ["pclass"],
x = predictors,
y = response,
training_frame = train,
validation_frame = valid)
# convert the segmented models to an H2OFrame
titanic_models.as_frame()