Merge pull request #5 from ottenbreit-data-science/bins3

Improved binning methodology

Author: mathias-von-ottenbreit

cpp/main.cpp

@@ -46,10 +46,10 @@ int main()
     //Saving results
     save_data("output.csv",predictions);
     std::cout<<"min validation_error "<<model.validation_error_steps.minCoeff()<<"\n\n";
-    std::cout<<check_if_approximately_equal(model.validation_error_steps.minCoeff(),6.95696,0.00001)<<"\n";
+    std::cout<<check_if_approximately_equal(model.validation_error_steps.minCoeff(),7.02559,0.00001)<<"\n";
 
     std::cout<<"mean prediction "<<predictions.mean()<<"\n\n";
-    std::cout<<check_if_approximately_equal(predictions.mean(),23.8927,0.0001)<<"\n";
+    std::cout<<check_if_approximately_equal(predictions.mean(),23.9213,0.0001)<<"\n";
 
     std::cout<<"best_m: "<<model.m<<"\n";
 

cpp/term.h

@@ -41,6 +41,7 @@ class Term
     VectorXd y_discretized;
     VectorXd errors_initial;
     double error_initial;
+    std::vector<size_t> observations_in_bins;
 
     //methods
     void calculate_error_where_given_terms_are_zero(const VectorXd &y, const VectorXd &sample_weight);
@@ -411,6 +412,13 @@ void Term::setup_bins()
         }
         bins_split_points_left.shrink_to_fit();
         bins_split_points_right.shrink_to_fit();
+
+        //observations in bins
+        observations_in_bins.reserve(bins_start_index.size());
+        for (size_t i = 0; i < bins_start_index.size(); ++i)
+        {
+            observations_in_bins.push_back(bins_end_index[i]-bins_start_index[i]+1);
+        }
     }
 }
 
@@ -421,22 +429,30 @@ void Term::discretize_data_by_bin()
         values_discretized.resize(bins_start_index.size());
         for (size_t i = 0; i < bins_start_index.size(); ++i)
         {
-            values_discretized[i]=sorted_vectors.values_sorted.block(bins_start_index[i],0,bins_end_index[i]-bins_start_index[i]+1,1).mean();
+            values_discretized[i]=sorted_vectors.values_sorted.block(bins_start_index[i],0,observations_in_bins[i],1).mean();
         }
 
-        if(sorted_vectors.sample_weight_sorted.size()>0)
+        sample_weight_discretized.resize(bins_start_index.size());
+        bool sample_weights_were_provided_by_user{sorted_vectors.sample_weight_sorted.size()>0};
+        if(sample_weights_were_provided_by_user)
         {
-            sample_weight_discretized.resize(bins_start_index.size());
             for (size_t i = 0; i < bins_start_index.size(); ++i)
             {
-                sample_weight_discretized[i]=sorted_vectors.sample_weight_sorted.block(bins_start_index[i],0,bins_end_index[i]-bins_start_index[i]+1,1).mean();
+                sample_weight_discretized[i]=sorted_vectors.sample_weight_sorted.block(bins_start_index[i],0,observations_in_bins[i],1).sum();
+            }
+        }
+        else
+        {
+            for (size_t i = 0; i < bins_start_index.size(); ++i)
+            {
+                sample_weight_discretized[i]=static_cast<double>(observations_in_bins[i]);
             }
         }
     }
     y_discretized.resize(bins_start_index.size());
     for (size_t i = 0; i < bins_start_index.size(); ++i)
     {
-        y_discretized[i]=sorted_vectors.y_sorted.block(bins_start_index[i],0,bins_end_index[i]-bins_start_index[i]+1,1).mean();
+        y_discretized[i]=sorted_vectors.y_sorted.block(bins_start_index[i],0,observations_in_bins[i],1).mean();
     }
 
     max_index_discretized=calculate_max_index_in_vector(values_discretized);
@@ -516,16 +532,8 @@ void Term::calculate_coefficient_and_error_on_discretized_data(bool direction_ri
     double xwy{0};
     for (size_t i = index_start; i <= index_end; ++i)
     {
-        if(sample_weight_discretized.size()>0)
-        {
-            xwx+=values_sorted[i]*values_sorted[i]*sample_weight_discretized[i];
-            xwy+=values_sorted[i]*y_discretized[i]*sample_weight_discretized[i];
-        }
-        else
-        {
-            xwx+=values_sorted[i]*values_sorted[i];
-            xwy+=values_sorted[i]*y_discretized[i];
-        }
+        xwx+=values_sorted[i]*values_sorted[i]*sample_weight_discretized[i];
+        xwy+=values_sorted[i]*y_discretized[i]*sample_weight_discretized[i];
     }
     if(xwx!=0)
     {
@@ -595,6 +603,7 @@ void Term::cleanup_after_fit()
     bins_end_index.clear();
     bins_split_points_left.clear();
     bins_split_points_right.clear();
+    observations_in_bins.clear();
     values_discretized.resize(0);
     sample_weight_discretized.resize(0);
 }

cpp/test ALRRegressor.cpp

@@ -16,7 +16,7 @@ int main()
     model.m=100;
     model.v=1.0;
     model.bins=10;
-    model.n_jobs=0;
+    model.n_jobs=1;
     model.loss_function_mse=true;
     model.verbosity=3;
     model.max_interaction_level=100;

setup.py

@@ -15,7 +15,7 @@
 
 setuptools.setup(
     name='aplr',
-    version='1.0.6',
+    version='1.0.7',
     description='Automatic Piecewise Linear Regression',
     ext_modules=[sfc_module],
     author="Mathias von Ottenbreit",