Package 'iBART'

The design matrix X containing all the primary features

A vector of unit of the primary features. For example, unit <- c("cm", "s"). Then the unit of $x1$ is centimeter and the unit of $x2$ is second.

A vector of dimension of the units. For example, unit <- c("cm", "s") and dimension <- c(2, 1) mean that the unit of $x1$ is square centimeter and the unit of $x2$ is second.

Input matrix of primary features $X$.

Response variable $y$.

Optional: name of primary features.

Optional: units and their respective dimensions of primary features. This is used to perform dimension analysis for generated descriptors to avoid generating unphyiscal descriptors, such as $size + size^2$. See generate_unit() for details.

Variable selection criterion used in BART. Three options are available: (1) "global_se", (2) "global_max", (3) "local". The default is "global_se". See var_selection_by_permute in R package bartMachine for more detail.

BART parameter: number of trees to be grown in the sum-of-trees model. If you want different values for each iteration of BART, input a vector of length equal to number of iterations. Default is num_trees = 20.

BART parameter: number of MCMC samples to be discarded as “burn-in". If you want different values for each iteration of BART, input a vector of length equal to number of iterations. Default is num_burn_in = 10000.

BART parameter: number of MCMC samples to draw from the posterior distribution of $hat{f}(x)$. If you want different values for each iteration of BART, input a vector of length equal to number of iterations. Default is num_iterations_after_burn_in = 5000.

BART parameter: number of replicates to over over to for the BART model's variable inclusion proportions. If you want different values for each iteration of BART, input a vector of length equal to number of iterations. Default is num_reps_for_avg = 10.

BART parameter: number of permutations of the response to be made to generate the “null” permutation distribution. If you want different values for each iteration of BART, input a vector of length equal to number of iterations. Default is num_permute_samples = 50.

glmnet parameter: loss to use for cross-validation. The default is type.measure="deviance", which uses squared-error for Gaussian models (a.k.a type.measure="mse" there). type.measure="mae" (mean absolute error) can be used also.

glmnet parameter: number of folds - default is 10. Smallest value allowable is nfolds=3.

glmnet parameter: the number of lambda values - default is 100.

glmnet parameter: If TRUE, then CV is done with respect to the mixing parameter gamma as well as lambda. Default is relax=FALSE.

glmnet parameter: the values of the parameter for mixing the relaxed fit with the regularized fit, between 0 and 1; default is gamma = c(0, 0.25, 0.5, 0.75, 1)

A vector of operation order. For example, opt = c("unary", "binary", "unary") will apply unary operators, then binary operators, then unary operators. Available operator sets are "unary", "binary", and "all", where "all" is the union of "unary" and "binary".

Logical flag for using $sin(\pi*x)$ and $cos(\pi*x)$ to generate descriptors. This is useful if you think there is periodic relationship between predictors and response. Default is sin_cos = FALSE.

Logical flag for applying non-negative-valued operators, such as $\sqrt x$ and $log(x)$, when some values of $x$ is negative. If apply_pos_opt_on_neg_x == TRUE, apply absolute value operator first then non-negative-valued operator, i.e. generate $\sqrt |x|$ and $log(|x|)$ instead. Default is apply_pos_opt_on_neg_x = TRUE.

Number of iterations to hold. This allows iBART to run consecutive operator transformations before screening. Note hold = 0 is equivalent to no skipping of variable selection in each iBART iterations. It should be less than iter.

Logical flag for pre-screening the primary features X using BART. Only selected primary features will be used to generate descriptors. Note that pre_screen = FALSE is equivalent to hold = 1.

Logical flag for screening out primary features that are independet of the response variable $y$.

Logical flag for out of sample assessment. Default is out_sample = FALSE.

Numerical vector storing the row indices for training data. Please set out_sample = TRUE if you supplied train_idx.

Proportion of data used to train model. Value must be between (0,1]. This is only needed when out_sample = TRUE and train_idx == NULL. Default is train_ratio = 1.

Logical flag for L-zero variable selection. Default is Lzero = TRUE.

Logical flag for parallel L-zero variable selection. Default is parallel = FALSE.

If Lzero == TRUE, K sets the maximum number of descriptors to be selected.

If Lzero == TRUE, logical flag for selecting best number of descriptors using AIC. Possible number of descriptors are $1 \le k \le K$.

Logical flag for data standardization prior to model fitting in BART and LASSO. Default is standardize = TRUE.

Logical flag for writing log file to working directory. The log file will contain information such as the descriptors selected by iBART, RMSE of the linear model build on the selected descriptors, etc. Default is writeLog = FALSE.

Logical flag for printing progress to console. Default is verbose = TRUE.

Internal parameter. Default is count = NULL.

Optional: sets the seed in both R and Java. Default is seed = NULL which does not set the seed in R nor Java.

The LASSO output of the last iteration of iBART. The predictors with non-zero coefficient are called the iBART selected descriptors.

The numerical values of the iBART selected descriptors.

The names of the iBART selected descriptors.

Coefficients of the iBART model. The first element is an intercept.

The training matrix used in the last iteration.

The testing matrix used in the last iteration.

The number of descriptors generated by iBART in each iteration.

The number of descriptors selected by iBART in each iteration.

In sample RMSE of the LASSO model.

Out of sample RMSE of the LASSO model if out_sample == TRUE.

The $l_0$-penalized regression models fitted on the iBART selected descriptors for $1 \le k \le K$.

The name of the best $k$D descriptors selected by the $l_0$-penalized regression model for $1 \le k \le K$.

In sample RMSE of the $l_0$-penalized regression model for $1 \le k \le K$.

Out of sample RMSE of the $l_0$-penalized regression model for $1 \le k \le K$ if out_sample == TRUE.

The best $l_0$-penalized regression model selected by AIC.

The best $k$D descriptors where $1 \le k \le K$ is chosen via AIC.

In sample RMSE of the best $l_0$-penalized regression models chosen by AIC.

Out of sample RMSE of the best $l_0$-penalized regression models chosen by AIC if out_sample == TRUE.

Runtime in second.

The design matrix used during training.

The response variable used during training.

The design matrix used during testing. Default is X_test = NULL and full data will be used to train the best subset linear regression model.

The response variable used during testing. Default is y_test = NULL and full data will be used to train the best subset linear regression model.

The maximum number of predictors allowed in the model. For example, k = 5 will produce the best model 5 predictors.

Logical flag for parallelization. Default is parallel = FALSE.

An lm object storing the best k-predictor linear model.

The variable name of the best k predictors.

In-sample RMSE of the model.

Out-of-sample RMSE of the model.