R/modelselection.R
h2o.modelSelection.Rd
H2O ModelSelection is used to build the best model with one predictor, two predictors, ... up to max_predictor_number specified in the algorithm parameters when mode=allsubsets. The best model is the one with the highest R2 value. When mode=maxr, the model returned is no longer guaranteed to have the best R2 value.
h2o.modelSelection( x, y, training_frame, model_id = NULL, validation_frame = NULL, nfolds = 0, seed = 1, fold_assignment = c("AUTO", "Random", "Modulo", "Stratified"), fold_column = NULL, ignore_const_cols = TRUE, score_each_iteration = FALSE, score_iteration_interval = 0, offset_column = NULL, weights_column = NULL, family = c("AUTO", "gaussian", "binomial", "fractionalbinomial", "quasibinomial", "poisson", "gamma", "tweedie", "negativebinomial"), link = c("family_default", "identity", "logit", "log", "inverse", "tweedie", "ologit"), tweedie_variance_power = 0, tweedie_link_power = 0, theta = 0, solver = c("AUTO", "IRLSM", "L_BFGS", "COORDINATE_DESCENT_NAIVE", "COORDINATE_DESCENT", "GRADIENT_DESCENT_LH", "GRADIENT_DESCENT_SQERR"), alpha = NULL, lambda = c(0), lambda_search = FALSE, early_stopping = FALSE, nlambdas = 0, standardize = TRUE, missing_values_handling = c("MeanImputation", "Skip", "PlugValues"), plug_values = NULL, compute_p_values = FALSE, remove_collinear_columns = FALSE, intercept = TRUE, non_negative = FALSE, max_iterations = 0, objective_epsilon = 1, beta_epsilon = 1e04, gradient_epsilon = 1, startval = NULL, prior = 0, cold_start = FALSE, lambda_min_ratio = 0, beta_constraints = NULL, max_active_predictors = 1, obj_reg = 1, stopping_rounds = 0, stopping_metric = c("AUTO", "deviance", "logloss", "MSE", "RMSE", "MAE", "RMSLE", "AUC", "AUCPR", "lift_top_group", "misclassification", "mean_per_class_error", "custom", "custom_increasing"), stopping_tolerance = 0.001, balance_classes = FALSE, class_sampling_factors = NULL, max_after_balance_size = 5, max_runtime_secs = 0, custom_metric_func = NULL, nparallelism = 0, max_predictor_number = 1, min_predictor_number = 1, mode = c("allsubsets", "maxr", "maxrsweep", "backward"), build_glm_model = FALSE, p_values_threshold = 0, influence = c("dfbetas"), multinode_mode = FALSE )
x  (Optional) A vector containing the names or indices of the predictor variables to use in building the model. If x is missing, then all columns except y are used. 

y  The name or column index of the response variable in the data. The response must be either a numeric or a categorical/factor variable. If the response is numeric, then a regression model will be trained, otherwise it will train a classification model. 
training_frame  Id of the training data frame. 
model_id  Destination id for this model; autogenerated if not specified. 
validation_frame  Id of the validation data frame. 
nfolds  Number of folds for Kfold crossvalidation (0 to disable or >= 2). Defaults to 0. 
seed  Seed for random numbers (affects certain parts of the algo that are stochastic and those might or might not be enabled by default). Defaults to 1 (timebased random number). 
fold_assignment  Crossvalidation fold assignment scheme, if fold_column is not specified. The 'Stratified' option will stratify the folds based on the response variable, for classification problems. Must be one of: "AUTO", "Random", "Modulo", "Stratified". Defaults to AUTO. 
fold_column  Column with crossvalidation fold index assignment per observation. 
ignore_const_cols 

score_each_iteration 

score_iteration_interval  Perform scoring for every score_iteration_interval iterations Defaults to 0. 
offset_column  Offset column. This will be added to the combination of columns before applying the link function. 
weights_column  Column with observation weights. Giving some observation a weight of zero is equivalent to excluding it from the dataset; giving an observation a relative weight of 2 is equivalent to repeating that row twice. Negative weights are not allowed. Note: Weights are perrow observation weights and do not increase the size of the data frame. This is typically the number of times a row is repeated, but noninteger values are supported as well. During training, rows with higher weights matter more, due to the larger loss function prefactor. If you set weight = 0 for a row, the returned prediction frame at that row is zero and this is incorrect. To get an accurate prediction, remove all rows with weight == 0. 
family  Family. For maxr/maxrsweep, only gaussian. For backward, ordinal and multinomial families are not supported Must be one of: "AUTO", "gaussian", "binomial", "fractionalbinomial", "quasibinomial", "poisson", "gamma", "tweedie", "negativebinomial". Defaults to AUTO. 
link  Link function. Must be one of: "family_default", "identity", "logit", "log", "inverse", "tweedie", "ologit". Defaults to family_default. 
tweedie_variance_power  Tweedie variance power Defaults to 0. 
tweedie_link_power  Tweedie link power Defaults to 0. 
theta  Theta Defaults to 0. 
solver  AUTO will set the solver based on given data and the other parameters. IRLSM is fast on on problems with small number of predictors and for lambdasearch with L1 penalty, L_BFGS scales better for datasets with many columns. Must be one of: "AUTO", "IRLSM", "L_BFGS", "COORDINATE_DESCENT_NAIVE", "COORDINATE_DESCENT", "GRADIENT_DESCENT_LH", "GRADIENT_DESCENT_SQERR". Defaults to IRLSM. 
alpha  Distribution of regularization between the L1 (Lasso) and L2 (Ridge) penalties. A value of 1 for alpha represents Lasso regression, a value of 0 produces Ridge regression, and anything in between specifies the amount of mixing between the two. Default value of alpha is 0 when SOLVER = 'LBFGS'; 0.5 otherwise. 
lambda  Regularization strength Defaults to c(0.0). 
lambda_search 

early_stopping 

nlambdas  Number of lambdas to be used in a search. Default indicates: If alpha is zero, with lambda search set to True, the value of nlamdas is set to 30 (fewer lambdas are needed for ridge regression) otherwise it is set to 100. Defaults to 0. 
standardize 

missing_values_handling  Handling of missing values. Either MeanImputation, Skip or PlugValues. Must be one of: "MeanImputation", "Skip", "PlugValues". Defaults to MeanImputation. 
plug_values  Plug Values (a single row frame containing values that will be used to impute missing values of the training/validation frame, use with conjunction missing_values_handling = PlugValues) 
compute_p_values 

remove_collinear_columns 

intercept 

non_negative 

max_iterations  Maximum number of iterations Defaults to 0. 
objective_epsilon  Converge if objective value changes less than this. Default (of 1.0) indicates: If lambda_search is set to True the value of objective_epsilon is set to .0001. If the lambda_search is set to False and lambda is equal to zero, the value of objective_epsilon is set to .000001, for any other value of lambda the default value of objective_epsilon is set to .0001. Defaults to 1. 
beta_epsilon  Converge if beta changes less (using Linfinity norm) than beta esilon, ONLY applies to IRLSM solver Defaults to 0.0001. 
gradient_epsilon  Converge if objective changes less (using Linfinity norm) than this, ONLY applies to LBFGS solver. Default (of 1.0) indicates: If lambda_search is set to False and lambda is equal to zero, the default value of gradient_epsilon is equal to .000001, otherwise the default value is .0001. If lambda_search is set to True, the conditional values above are 1E8 and 1E6 respectively. Defaults to 1. 
startval  double array to initialize fixed and random coefficients for HGLM, coefficients for GLM. 
prior  Prior probability for y==1. To be used only for logistic regression iff the data has been sampled and the mean of response does not reflect reality. Defaults to 0. 
cold_start 

lambda_min_ratio  Minimum lambda used in lambda search, specified as a ratio of lambda_max (the smallest lambda that drives all coefficients to zero). Default indicates: if the number of observations is greater than the number of variables, then lambda_min_ratio is set to 0.0001; if the number of observations is less than the number of variables, then lambda_min_ratio is set to 0.01. Defaults to 0. 
beta_constraints  Beta constraints 
max_active_predictors  Maximum number of active predictors during computation. Use as a stopping criterion to prevent expensive model building with many predictors. Default indicates: If the IRLSM solver is used, the value of max_active_predictors is set to 5000 otherwise it is set to 100000000. Defaults to 1. 
obj_reg  Likelihood divider in objective value computation, default (of 1.0) will set it to 1/nobs Defaults to 1. 
stopping_rounds  Early stopping based on convergence of stopping_metric. Stop if simple moving average of length k of the stopping_metric does not improve for k:=stopping_rounds scoring events (0 to disable) Defaults to 0. 
stopping_metric  Metric to use for early stopping (AUTO: logloss for classification, deviance for regression and anomaly_score for Isolation Forest). Note that custom and custom_increasing can only be used in GBM and DRF with the Python client. Must be one of: "AUTO", "deviance", "logloss", "MSE", "RMSE", "MAE", "RMSLE", "AUC", "AUCPR", "lift_top_group", "misclassification", "mean_per_class_error", "custom", "custom_increasing". Defaults to AUTO. 
stopping_tolerance  Relative tolerance for metricbased stopping criterion (stop if relative improvement is not at least this much) Defaults to 0.001. 
balance_classes 

class_sampling_factors  Desired over/undersampling ratios per class (in lexicographic order). If not specified, sampling factors will be automatically computed to obtain class balance during training. Requires balance_classes. 
max_after_balance_size  Maximum relative size of the training data after balancing class counts (can be less than 1.0). Requires balance_classes. Defaults to 5.0. 
max_runtime_secs  Maximum allowed runtime in seconds for model training. Use 0 to disable. Defaults to 0. 
custom_metric_func  Reference to custom evaluation function, format: `language:keyName=funcName` 
nparallelism  number of models to build in parallel. Defaults to 0.0 which is adaptive to the system capability Defaults to 0. 
max_predictor_number  Maximum number of predictors to be considered when building GLM models. Defaults to 1. Defaults to 1. 
min_predictor_number  For mode = 'backward' only. Minimum number of predictors to be considered when building GLM models starting with all predictors to be included. Defaults to 1. Defaults to 1. 
mode  Mode: Used to choose model selection algorithms to use. Options include 'allsubsets' for all subsets, 'maxr' that uses sequential replacement and GLM to build all models, slow but works with crossvalidation, validation frames for more robust results, 'maxrsweep' that uses sequential replacement and sweeping action, much faster than 'maxr', 'backward' for backward selection. Must be one of: "allsubsets", "maxr", "maxrsweep", "backward". Defaults to maxr. 
build_glm_model 

p_values_threshold  For mode='backward' only. If specified, will stop the model building process when all coefficientspvalues drop below this threshold Defaults to 0. 
influence  If set to dfbetas will calculate the difference in beta when a datarow is included and excluded in the dataset. Must be one of: "dfbetas". 
multinode_mode 

if (FALSE) { library(h2o) h2o.init() # Run ModelSelection of VOL ~ all predictors prostate_path < system.file("extdata", "prostate.csv", package = "h2o") prostate < h2o.uploadFile(path = prostate_path) prostate$CAPSULE < as.factor(prostate$CAPSULE) model < h2o.modelSelection(y="VOL", x=c("RACE","AGE","RACE","DPROS"), training_frame=prostate) }