Changeset 1182 for trunk/yat

Timestamp:

Feb 28, 2008, 1:27:37 PM (14 years ago)

Author:

Peter

Message:

working on #335. Fixed weighted test data case. Left to fix is when there are missing features in training in other words what should happen when complete training cannot be done because lack of data. The current behavior is probably not optimal, but have to look into it in more detail.

Location:

trunk/yat/classifier

Files:

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

trunk/yat/classifier/NBC.cc

@@ -152 +152 @@
     prediction.resize(centroids_.columns(), mlw.columns(), 0);
 
-    // first calculate -lnP = sum sigma_i + (x_i-m_i)^2/2sigma_i^2
+    // first calculate -lnP = sum (sigma_i) + 
+    // N sum w_i(x_i-m_i)^2/2sigma_i^2 / sum w_i
     for (size_t label=0; label<centroids_.columns(); ++label) {
       double sum_log_sigma = sum_logsigma(label);
       for (size_t sample=0; sample<prediction.rows(); ++sample) {
-        prediction(label,sample) = sum_log_sigma;
+        statistics::AveragerWeighted aw;
         for (size_t i=0; i<mlw.rows(); ++i) 
-          // 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)/
-              sigma2_(i, label);
-          }
-        
+          // missing training features
+          if (!std::isnan(sigma2_(i, label))) 
+            aw.add(std::pow(mlw.data(i, label)-centroids_(i, label),2)/
+                   sigma2_(i, label), mlw.weight(i, label));
+        prediction(label,sample) = sum_log_sigma + mlw.rows()*aw.mean()/2;
       }
     }

trunk/yat/classifier/NBC.h

@@ -104 +104 @@
        row in \a res corresponds to a class. The prediction is the
        estimated probability that sample belong to class \f$ j \f$
+
+       \f$ P_j = \frac{1}{Z}\prod_i\({\frac{1}{\sqrt{2\pi\sigma_i^2}}}\)
+       \exp(\frac{\sum{w_i(x_i-\mu_i)^2}{\sigma_i^2}}{\sum w_i})\f$, 
+       where \f$ \mu_i
+       \f$ and \f$ \sigma_i^2 \f$ are the estimated mean and variance,
+       respectively. If a \f$ \sigma_i \f$ could not be estimated
+       during training, corresponding factor is set to unity, in other
+       words, that feature is ignored for the prediction of that
+       particular class. Z is chosen such that total probability, \f$
+       \sum P_j \f$, equals unity. 
      */
     void predict(const MatrixLookupWeighted& data, utility::Matrix& res) const;