Merge pull request #181 from StochasticTree/bcf-feature-subsets-hotfix

Hotfix for two issues with GFR feature subset sampling in BCF

Author: andrewherren

R/bcf.R

@@ -503,17 +503,6 @@ bcf <- function(X_train, Z_train, y_train, propensity_train = NULL, rfx_group_id
     X_test_raw <- X_test
     if (!is.null(X_test)) X_test <- preprocessPredictionData(X_test, X_train_metadata)
 
-    # Set num_features_subsample to default, ncol(X_train), if not already set
-    if (is.null(num_features_subsample_mu)) {
-        num_features_subsample_mu <- ncol(X_train)
-    }
-    if (is.null(num_features_subsample_tau)) {
-        num_features_subsample_tau <- ncol(X_train)
-    }
-    if (is.null(num_features_subsample_variance)) {
-        num_features_subsample_variance <- ncol(X_train)
-    }
-    
     # Convert all input data to matrices if not already converted
     if ((is.null(dim(Z_train))) && (!is.null(Z_train))) {
         Z_train <- as.matrix(as.numeric(Z_train))
@@ -722,6 +711,17 @@ bcf <- function(X_train, Z_train, y_train, propensity_train = NULL, rfx_group_id
         variable_weights_variance <- variable_weights_variance / sum(variable_weights_variance)
     }
 
+    # Set num_features_subsample to default, ncol(X_train), if not already set
+    if (is.null(num_features_subsample_mu)) {
+        num_features_subsample_mu <- ncol(X_train)
+    }
+    if (is.null(num_features_subsample_tau)) {
+        num_features_subsample_tau <- ncol(X_train)
+    }
+    if (is.null(num_features_subsample_variance)) {
+        num_features_subsample_variance <- ncol(X_train)
+    }
+    
     # Preliminary runtime checks for probit link
     if (probit_outcome_model) {
         if (!(length(unique(y_train)) == 2)) {

include/stochtree/tree_sampler.h

@@ -716,18 +716,28 @@ static inline void GFRSampleTreeOneIter(Tree* tree, ForestTracker& tracker, Fore
   // Subsample features (if requested)
   std::vector<bool> feature_subset(p, true);
   if (num_features_subsample < p) {
-    std::vector<int> feature_indices(p);
-    std::iota(feature_indices.begin(), feature_indices.end(), 0);
-    std::vector<int> features_selected(num_features_subsample);
-    sample_without_replacement<int, double>(
-      features_selected.data(), variable_weights.data(), feature_indices.data(), 
-      p, num_features_subsample, gen
-    );
-    for (int i = 0; i < p; i++) {
-      feature_subset.at(i) = false;
+    // Check if the number of (meaningfully) nonzero selection probabilities is greater than num_features_subsample
+    int number_nonzero_weights = 0;
+    for (int j = 0; j < p; j++) {
+      if (std::abs(variable_weights.at(j)) > kEpsilon) {
+        number_nonzero_weights++;
+      }
     }
-    for (const auto& feat : features_selected) {
-      feature_subset.at(feat) = true;
+    if (number_nonzero_weights > num_features_subsample) {
+      // Sample with replacement according to variable_weights
+      std::vector<int> feature_indices(p);
+      std::iota(feature_indices.begin(), feature_indices.end(), 0);
+      std::vector<int> features_selected(num_features_subsample);
+      sample_without_replacement<int, double>(
+        features_selected.data(), variable_weights.data(), feature_indices.data(), 
+        p, num_features_subsample, gen
+      );
+      for (int i = 0; i < p; i++) {
+        feature_subset.at(i) = false;
+      }
+      for (const auto& feat : features_selected) {
+        feature_subset.at(feat) = true;
+      }
     }
   }
 
@@ -782,6 +792,7 @@ static inline void GFRSampleTreeOneIter(Tree* tree, ForestTracker& tracker, Fore
  * \param pre_initialized Whether or not `active_forest` has already been initialized (note: this parameter will be refactored out soon).
  * \param backfitting Whether or not the sampler uses "backfitting" (wherein the sampler for a given tree only depends on the other trees via
  * their effect on the residual) or the more general "blocked MCMC" (wherein the state of other trees must be more explicitly considered).
+ * \param num_features_subsample How many features to subsample when running the GFR algorithm.
  * \param leaf_suff_stat_args Any arguments which must be supplied to initialize a `LeafSuffStat` object.
  */
 template <typename LeafModel, typename LeafSuffStat, typename... LeafSuffStatConstructorArgs>

stochtree/bcf.py

@@ -1093,14 +1093,6 @@ def sample(
         else:
             variable_subset_variance = [i for i in range(X_train.shape[1])]
 
-        # Set num_features_subsample to default, ncol(X_train), if not already set
-        if num_features_subsample_mu is None:
-            num_features_subsample_mu = X_train.shape[1]
-        if num_features_subsample_tau is None:
-            num_features_subsample_tau = X_train.shape[1]
-        if num_features_subsample_variance is None:
-            num_features_subsample_variance = X_train.shape[1]
-
         # Determine whether a test set is provided
         self.has_test = X_test is not None
 
@@ -1498,6 +1490,14 @@ def sample(
         # Store propensity score requirements of the BCF forests
         self.propensity_covariate = propensity_covariate
 
+        # Set num_features_subsample to default, ncol(X_train), if not already set
+        if num_features_subsample_mu is None:
+            num_features_subsample_mu = X_train_processed.shape[1]
+        if num_features_subsample_tau is None:
+            num_features_subsample_tau = X_train_processed.shape[1]
+        if num_features_subsample_variance is None:
+            num_features_subsample_variance = X_train_processed.shape[1]
+
         # Container of variance parameter samples
         self.num_gfr = num_gfr
         self.num_burnin = num_burnin

vignettes/CausalInference.Rmd

@@ -949,7 +949,7 @@ sqrt(mean((test_outcome_mean - y_test)^2))
 
 #### MCMC, covariate subset in $\tau(X)$
 
-Here we simulate from the model with the original MCMC sampler, using only covariate $X_1$ in the treatment effect forest.
+Here we simulate from the model with the original MCMC sampler, using only covariates $X_1$ and $X_2$ in the treatment effect forest.
 
 ```{r}
 num_gfr <- 0
@@ -1073,7 +1073,7 @@ sqrt(mean((y_test - test_outcome_mean)^2))
 
 #### Warmstart, covariate subset in $\tau(X)$
 
-Here we simulate from the model with the warm-start sampler, using only covariate $X_1$ in the treatment effect forest.
+Here we simulate from the model with the warm-start sampler, using only covariates $X_1$ and $X_2$ in the treatment effect forest.
 
 ```{r}
 num_gfr <- 10