Merge pull request #52 from StochasticTree/bcf-feature-subsets-mu-tau

Using subsets of features in the prognostic and treatment forests in BCF

Author: andrewherren

R/bart.R

@@ -33,9 +33,10 @@
 #' @param nu Shape parameter in the `IG(nu, nu*lambda)` global error variance model. Default: 3.
 #' @param lambda Component of the scale parameter in the `IG(nu, nu*lambda)` global error variance prior. If not specified, this is calibrated as in Sparapani et al (2021).
 #' @param a_leaf Shape parameter in the `IG(a_leaf, b_leaf)` leaf node parameter variance model. Default: 3.
-#' @param b_leaf Scale parameter in the `IG(a_leaf, b_leaf)` leaf node parameter variance model. Calibrated internally as 0.5/num_trees if not set here.
+#' @param b_leaf Scale parameter in the `IG(a_leaf, b_leaf)` leaf node parameter variance model. Calibrated internally as `0.5/num_trees` if not set here.
 #' @param q Quantile used to calibrated `lambda` as in Sparapani et al (2021). Default: 0.9.
 #' @param sigma2_init Starting value of global variance parameter. Calibrated internally as in Sparapani et al (2021) if not set here.
+#' @param variable_weights Numeric weights reflecting the relative probability of splitting on each variable. Does not need to sum to 1 but cannot be negative. Defaults to `rep(1/ncol(X_train), ncol(X_train))` if not set here.
 #' @param num_trees Number of trees in the ensemble. Default: 200.
 #' @param num_gfr Number of "warm-start" iterations run using the grow-from-root algorithm (He and Hahn, 2021). Default: 5.
 #' @param num_burnin Number of "burn-in" iterations of the MCMC sampler. Default: 0.
@@ -80,10 +81,19 @@ bart <- function(X_train, y_train, W_train = NULL, group_ids_train = NULL,
                  cutpoint_grid_size = 100, tau_init = NULL, alpha = 0.95, 
                  beta = 2.0, min_samples_leaf = 5, leaf_model = 0, 
                  nu = 3, lambda = NULL, a_leaf = 3, b_leaf = NULL, 
-                 q = 0.9, sigma2_init = NULL, num_trees = 200, num_gfr = 5, 
-                 num_burnin = 0, num_mcmc = 100, sample_sigma = T, 
-                 sample_tau = T, random_seed = -1, keep_burnin = F, 
-                 keep_gfr = F, verbose = F){
+                 q = 0.9, sigma2_init = NULL, variable_weights = NULL, 
+                 num_trees = 200, num_gfr = 5, num_burnin = 0, 
+                 num_mcmc = 100, sample_sigma = T, sample_tau = T, 
+                 random_seed = -1, keep_burnin = F, keep_gfr = F, 
+                 verbose = F){
+    # Variable weight preprocessing (and initialization if necessary)
+    if (is.null(variable_weights)) {
+        variable_weights = rep(1/ncol(X_train), ncol(X_train))
+    }
+    if (any(variable_weights < 0)) {
+        stop("variable_weights cannot have any negative weights")
+    }
+    
     # Preprocess covariates
     if ((!is.data.frame(X_train)) && (!is.matrix(X_train))) {
         stop("X_train must be a matrix or dataframe")
@@ -93,12 +103,20 @@ bart <- function(X_train, y_train, W_train = NULL, group_ids_train = NULL,
             stop("X_test must be a matrix or dataframe")
         }
     }
+    if (ncol(X_train) != length(variable_weights)) {
+        stop("length(variable_weights) must equal ncol(X_train)")
+    }
     train_cov_preprocess_list <- preprocessTrainData(X_train)
     X_train_metadata <- train_cov_preprocess_list$metadata
     X_train <- train_cov_preprocess_list$data
+    original_var_indices <- X_train_metadata$original_var_indices
     feature_types <- X_train_metadata$feature_types
     if (!is.null(X_test)) X_test <- preprocessPredictionData(X_test, X_train_metadata)
 
+    # Update variable weights
+    variable_weights_adj <- 1/sapply(original_var_indices, function(x) sum(original_var_indices == x))
+    variable_weights <- variable_weights[original_var_indices]*variable_weights_adj
+    
     # Convert all input data to matrices if not already converted
     if ((is.null(dim(W_train))) && (!is.null(W_train))) {
         W_train <- as.matrix(W_train)
@@ -295,9 +313,6 @@ bart <- function(X_train, y_train, W_train = NULL, group_ids_train = NULL,
     if (sample_sigma) global_var_samples <- rep(0, num_samples)
     if (sample_tau) leaf_scale_samples <- rep(0, num_samples)
 
-    # Variable selection weights
-    variable_weights <- rep(1/ncol(X_train), ncol(X_train))
-
     # Run GFR (warm start) if specified
     if (num_gfr > 0){
         gfr_indices = 1:num_gfr

R/bcf.R

@@ -2,6 +2,7 @@
 #' types. Matrices will be passed through assuming all columns are numeric.
 #'
 #' @param input_data Covariates, provided as either a dataframe or a matrix
+#' @param variable_weights Numeric weights reflecting the relative probability of splitting on each variable
 #'
 #' @return List with preprocessed (unmodified) data and details on the number of each type 
 #' of variable, unique categories associated with categorical variables, and the 
@@ -63,6 +64,7 @@ preprocessPredictionData <- function(input_data, metadata) {
 #' Returns a list including a matrix of preprocessed covariate values and associated tracking.
 #'
 #' @param input_matrix Covariate matrix.
+#' @param variable_weights Numeric weights reflecting the relative probability of splitting on each variable
 #'
 #' @return List with preprocessed (unmodified) data and details on the number of each type 
 #' of variable, unique categories associated with categorical variables, and the 
@@ -97,7 +99,8 @@ preprocessTrainMatrix <- function(input_matrix) {
         num_ordered_cat_vars = num_ordered_cat_vars, 
         num_unordered_cat_vars = num_unordered_cat_vars, 
         num_numeric_vars = num_numeric_vars, 
-        numeric_vars = numeric_vars
+        numeric_vars = numeric_vars, 
+        original_var_indices = 1:num_numeric_vars
     )
     output <- list(
         data = X, 
@@ -139,6 +142,7 @@ preprocessPredictionMatrix <- function(input_matrix, metadata) {
 #'
 #' @param input_df Dataframe of covariates. Users must pre-process any 
 #' categorical variables as factors (ordered for ordered categorical).
+#' @param variable_weights Numeric weights reflecting the relative probability of splitting on each variable
 #'
 #' @return List with preprocessed data and details on the number of each type 
 #' of variable, unique categories associated with categorical variables, and the 
@@ -164,6 +168,7 @@ preprocessTrainDataFrame <- function(input_df) {
     ordered_mask <- sapply(input_df, is.ordered)
     ordered_cat_matches <- factor_mask & ordered_mask
     ordered_cat_vars <- df_vars[ordered_cat_matches]
+    ordered_cat_var_inds <- unname(which(ordered_cat_matches))
     num_ordered_cat_vars <- length(ordered_cat_vars)
     if (num_ordered_cat_vars > 0) ordered_cat_df <- input_df[,ordered_cat_vars,drop=F]
 
@@ -173,12 +178,14 @@ preprocessTrainDataFrame <- function(input_df) {
     character_mask <- sapply(input_df, is.character)
     unordered_cat_matches <- (factor_mask & (!ordered_mask)) | character_mask
     unordered_cat_vars <- df_vars[unordered_cat_matches]
+    unordered_cat_var_inds <- unname(which(unordered_cat_matches))
     num_unordered_cat_vars <- length(unordered_cat_vars)
     if (num_unordered_cat_vars > 0) unordered_cat_df <- input_df[,unordered_cat_vars,drop=F]
 
     # Numeric variables
     numeric_matches <- (!ordered_cat_matches) & (!unordered_cat_matches)
     numeric_vars <- df_vars[numeric_matches]
+    numeric_var_inds <- unname(which(numeric_matches))
     num_numeric_vars <- length(numeric_vars)
     if (num_numeric_vars > 0) numeric_df <- input_df[,numeric_vars,drop=F]
 
@@ -187,6 +194,7 @@ preprocessTrainDataFrame <- function(input_df) {
     unordered_unique_levels <- list()
     ordered_unique_levels <- list()
     feature_types <- integer(0)
+    original_var_indices <- integer(0)
 
     # First, extract the numeric covariates
     if (num_numeric_vars > 0) {
@@ -197,6 +205,7 @@ preprocessTrainDataFrame <- function(input_df) {
         }
         X <- cbind(X, unname(Xnum))
         feature_types <- c(feature_types, rep(0, ncol(Xnum)))
+        original_var_indices <- c(original_var_indices, numeric_var_inds)
     }
 
     # Next, run some simple preprocessing on the ordered categorical covariates
@@ -210,6 +219,7 @@ preprocessTrainDataFrame <- function(input_df) {
         }
         X <- cbind(X, unname(Xordcat))
         feature_types <- c(feature_types, rep(1, ncol(Xordcat)))
+        original_var_indices <- c(original_var_indices, ordered_cat_var_inds)
     }
 
     # Finally, one-hot encode the unordered categorical covariates
@@ -220,6 +230,8 @@ preprocessTrainDataFrame <- function(input_df) {
             encode_list <- oneHotInitializeAndEncode(unordered_cat_df[,i])
             unordered_unique_levels[[var_name]] <- encode_list$unique_levels
             one_hot_mats[[var_name]] <- encode_list$Xtilde
+            one_hot_var <- rep(unordered_cat_var_inds[i], ncol(encode_list$Xtilde))
+            original_var_indices <- c(original_var_indices, one_hot_var)
         }
         Xcat <- do.call(cbind, one_hot_mats)
         X <- cbind(X, unname(Xcat))
@@ -231,7 +243,8 @@ preprocessTrainDataFrame <- function(input_df) {
         feature_types = feature_types, 
         num_ordered_cat_vars = num_ordered_cat_vars, 
         num_unordered_cat_vars = num_unordered_cat_vars, 
-        num_numeric_vars = num_numeric_vars
+        num_numeric_vars = num_numeric_vars, 
+        original_var_indices = original_var_indices
     )
     if (num_ordered_cat_vars > 0) {
         metadata[["ordered_cat_vars"]] = ordered_cat_vars

R/utils.R

@@ -93,7 +93,7 @@ pip install matplotlib seaborn jupyterlab
 The package can be installed in R via
 
 ```
-remotes::install_github("StochasticTree/stochtree-cpp")
+remotes::install_github("StochasticTree/stochtree-cpp", ref="r-dev")
 ```
 
 # C++ Core

README.md

@@ -81,12 +81,11 @@ class TreeEnsemble {
 
   inline void PredictInplace(ForestDataset& dataset, std::vector<double> &output, 
                              int tree_begin, int tree_end, data_size_t offset = 0) {
-    if (dataset.HasBasis()) {
-      CHECK(!is_leaf_constant_);
-      PredictInplace(dataset.GetCovariates(), dataset.GetBasis(), output, tree_begin, tree_end, offset);
-    } else {
-      CHECK(is_leaf_constant_);
+    if (is_leaf_constant_) {
       PredictInplace(dataset.GetCovariates(), output, tree_begin, tree_end, offset);
+    } else {
+      CHECK(dataset.HasBasis());
+      PredictInplace(dataset.GetCovariates(), dataset.GetBasis(), output, tree_begin, tree_end, offset);
     }
   }
 

include/stochtree/ensemble.h

@@ -71,7 +71,8 @@ class GaussianConstantLeafModel {
   std::tuple<double, double, data_size_t, data_size_t> EvaluateExistingSplit(ForestDataset& dataset, ForestTracker& tracker, ColumnVector& residual, double global_variance, int tree_num, int split_node_id, int left_node_id, int right_node_id);
   void EvaluateAllPossibleSplits(ForestDataset& dataset, ForestTracker& tracker, ColumnVector& residual, TreePrior& tree_prior, double global_variance, int tree_num, int split_node_id, 
                                  std::vector<double>& log_cutpoint_evaluations, std::vector<int>& cutpoint_features, std::vector<double>& cutpoint_values, std::vector<FeatureType>& cutpoint_feature_types, 
-                                 data_size_t& valid_cutpoint_count, CutpointGridContainer& cutpoint_grid_container, data_size_t node_begin, data_size_t node_end, std::vector<FeatureType>& feature_types);
+                                 data_size_t& valid_cutpoint_count, CutpointGridContainer& cutpoint_grid_container, data_size_t node_begin, data_size_t node_end, std::vector<double>& variable_weights, 
+                                 std::vector<FeatureType>& feature_types);
   double SplitLogMarginalLikelihood(GaussianConstantSuffStat& left_stat, GaussianConstantSuffStat& right_stat, double global_variance);
   double NoSplitLogMarginalLikelihood(GaussianConstantSuffStat& suff_stat, double global_variance);
   double PosteriorParameterMean(GaussianConstantSuffStat& suff_stat, double global_variance);
@@ -136,7 +137,8 @@ class GaussianUnivariateRegressionLeafModel {
   std::tuple<double, double, data_size_t, data_size_t> EvaluateExistingSplit(ForestDataset& dataset, ForestTracker& tracker, ColumnVector& residual, double global_variance, int tree_num, int split_node_id, int left_node_id, int right_node_id);
   void EvaluateAllPossibleSplits(ForestDataset& dataset, ForestTracker& tracker, ColumnVector& residual, TreePrior& tree_prior, double global_variance, int tree_num, int split_node_id, 
                                  std::vector<double>& log_cutpoint_evaluations, std::vector<int>& cutpoint_features, std::vector<double>& cutpoint_values, std::vector<FeatureType>& cutpoint_feature_types, 
-                                 data_size_t& valid_cutpoint_count, CutpointGridContainer& cutpoint_grid_container, data_size_t node_begin, data_size_t node_end, std::vector<FeatureType>& feature_types);
+                                 data_size_t& valid_cutpoint_count, CutpointGridContainer& cutpoint_grid_container, data_size_t node_begin, data_size_t node_end, std::vector<double>& variable_weights, 
+                                 std::vector<FeatureType>& feature_types);
   double SplitLogMarginalLikelihood(GaussianUnivariateRegressionSuffStat& left_stat, GaussianUnivariateRegressionSuffStat& right_stat, double global_variance);
   double NoSplitLogMarginalLikelihood(GaussianUnivariateRegressionSuffStat& suff_stat, double global_variance);
   double PosteriorParameterMean(GaussianUnivariateRegressionSuffStat& suff_stat, double global_variance);
@@ -203,7 +205,8 @@ class GaussianMultivariateRegressionLeafModel {
   std::tuple<double, double, data_size_t, data_size_t> EvaluateExistingSplit(ForestDataset& dataset, ForestTracker& tracker, ColumnVector& residual, double global_variance, int tree_num, int split_node_id, int left_node_id, int right_node_id);
   void EvaluateAllPossibleSplits(ForestDataset& dataset, ForestTracker& tracker, ColumnVector& residual, TreePrior& tree_prior, double global_variance, int tree_num, int split_node_id, 
                                  std::vector<double>& log_cutpoint_evaluations, std::vector<int>& cutpoint_features, std::vector<double>& cutpoint_values, std::vector<FeatureType>& cutpoint_feature_types, 
-                                 data_size_t& valid_cutpoint_count, CutpointGridContainer& cutpoint_grid_container, data_size_t node_begin, data_size_t node_end, std::vector<FeatureType>& feature_types);
+                                 data_size_t& valid_cutpoint_count, CutpointGridContainer& cutpoint_grid_container, data_size_t node_begin, data_size_t node_end, std::vector<double>& variable_weights, 
+                                 std::vector<FeatureType>& feature_types);
   double SplitLogMarginalLikelihood(GaussianMultivariateRegressionSuffStat& left_stat, GaussianMultivariateRegressionSuffStat& right_stat, double global_variance);
   double NoSplitLogMarginalLikelihood(GaussianMultivariateRegressionSuffStat& suff_stat, double global_variance);
   Eigen::VectorXd PosteriorParameterMean(GaussianMultivariateRegressionSuffStat& suff_stat, double global_variance);