Changeset 960 for trunk/yat/classifier

Timestamp:

Oct 10, 2007, 7:44:31 PM (16 years ago)

Author:

Peter

Message:

fixed ticket:271

Location:

trunk/yat/classifier

Files:

• NBC.cc (modified) (4 diffs)
• NBC.h (modified) (1 diff)

trunk/yat/classifier/NBC.cc

@@ -87 +87 @@
     utility::matrix nof_in_class(data_.rows(), target_.nof_classes());
     
-    for(size_t i=0; i<data_.rows(); ++i) {
-      std::vector<statistics::AveragerWeighted> aver(target_.nof_classes());
-      for(size_t j=0; j<data_.columns(); ++j) {
-        if (data_.weighted()){
-          const MatrixLookupWeighted& data = 
-            dynamic_cast<const MatrixLookupWeighted&>(data_);
+    // unweighted
+    if (data_.weighted()){
+      const MatrixLookupWeighted& data = 
+        dynamic_cast<const MatrixLookupWeighted&>(data_);
+      for(size_t i=0; i<data_.rows(); ++i) {
+        std::vector<statistics::AveragerWeighted> aver(target_.nof_classes());
+        for(size_t j=0; j<data_.columns(); ++j) 
           aver[target_(j)].add(data.data(i,j), data.weight(i,j));
-        }
-        else
-          aver[target_(j)].add(data_(i,j),1.0);
-      }
-      assert(centroids_.columns()==target_.nof_classes());
-      for (size_t j=0; j<target_.nof_classes(); ++j){
-        assert(i<centroids_.rows());
-        assert(j<centroids_.columns());
-        centroids_(i,j) = aver[j].mean();
-        assert(i<sigma2_.rows());
-        assert(j<sigma2_.columns());
-        sigma2_(i,j) = aver[j].variance();
+
+        assert(centroids_.columns()==target_.nof_classes());
+        for (size_t j=0; j<target_.nof_classes(); ++j){
+          assert(i<centroids_.rows());
+          assert(j<centroids_.columns());
+          centroids_(i,j) = aver[j].mean();
+          assert(i<sigma2_.rows());
+          assert(j<sigma2_.columns());
+          if (aver[j].variance())
+            sigma2_(i,j) = aver[j].variance();
+          else 
+            sigma2_(i,j) = std::numeric_limits<double>::quiet_NaN();
+        }
+      }
+    }
+    else {  
+      const MatrixLookup& data = dynamic_cast<const MatrixLookup&>(data_);
+      for(size_t i=0; i<data_.rows(); ++i) {
+        std::vector<statistics::Averager> aver(target_.nof_classes());
+        for(size_t j=0; j<data_.columns(); ++j) 
+          aver[target_(j)].add(data(i,j));
+
+        assert(centroids_.columns()==target_.nof_classes());
+        for (size_t j=0; j<target_.nof_classes(); ++j){
+          assert(i<centroids_.rows());
+          assert(j<centroids_.columns());
+          centroids_(i,j) = aver[j].mean();
+          assert(i<sigma2_.rows());
+          assert(j<sigma2_.columns());
+          if (aver[j].variance())
+            sigma2_(i,j) = aver[j].variance();
+          else 
+            sigma2_(i,j) = std::numeric_limits<double>::quiet_NaN();
+        }
       }
     }   
@@ -133 +156 @@
           prediction(label,sample) = sum_log_sigma;
           for (size_t i=0; i<x.rows(); ++i) 
-            // taking care of NaN
-            if (mlw->weight(i, label)) {
+            // taking care of NaN and missing training features
+            if (mlw->weight(i, label) && !std::isnan(sigma2_(i, label))) {
               prediction(label, sample) += mlw->weight(i, label)*
                 std::pow(mlw->data(i, label)-centroids_(i, label),2)/
@@ -151 +174 @@
           prediction(label,sample) = sum_log_sigma;
           for (size_t i=0; i<ml->rows(); ++i) 
-            prediction(label, sample) += 
-              std::pow((*ml)(i, label)-centroids_(i, label),2)/sigma2_(i, label);
+            // Ignoring missing features
+            if (!std::isnan(sigma2_(i, label)))
+              prediction(label, sample) += 
+                std::pow((*ml)(i, label)-centroids_(i, label),2)/
+                sigma2_(i, label);
         }
       }
@@ -190 +216 @@
     assert(label<sigma2_.columns());
     for (size_t i=0; i<sigma2_.rows(); ++i) {
-      sum_log_sigma += std::log(sigma2_(i, label));
+      if (!std::isnan(sigma2_(i,label))) 
+        sum_log_sigma += std::log(sigma2_(i, label));
     }
     return sum_log_sigma / 2; // taking sum of log(sigma) not sigma2

trunk/yat/classifier/NBC.h

@@ -78 +78 @@
     /// feature (see Averager and AveragerWeighted for details).
     ///
+    /// If variance can not be estimated (too few data points or all
+    /// points identical) for a feature and label, then that feature
+    /// is ignored for that specific label.
+    ///
     /// @return true if training succedeed.
     ///