#3 correct variance and covariance usage in preparation for a replacement by another algorithms

Author: mbok

src/main/java/org/scaleborn/linereg/calculation/statistics/StatsCalculator.java

@@ -30,40 +30,40 @@ public Statistics calculate(final StatsModel model) {
     final double[] slopeCoefficients = model.getSlopeCoefficients().getCoefficients();
 
     final double responseVariance = model.getStatsSampling().getResponseVariance();
-    double squaredError = responseVariance;
+    double squaredMeanError = responseVariance;
 
     for (int i = 0; i < featuresCount; i++) {
       final double c = slopeCoefficients[i];
       final double c2 = c * c;
       // Minus double of feature response coefficient
-      squaredError -= 2 * featuresResponseCovariance[i] * c;
+      squaredMeanError -= 2 * featuresResponseCovariance[i] * c;
 
       // Add values from covariance matrix of the derivation matrix
       for (int j = 0; j <= i; j++) {
         if (i == j) {
           // Variance term
-          squaredError += c2 * covarianceLowerTriangularMatrix[i][j];
+          squaredMeanError += c2 * covarianceLowerTriangularMatrix[i][j];
         } else {
           // Covariance term
-          squaredError += 2 * c * slopeCoefficients[j] * covarianceLowerTriangularMatrix[i][j];
+          squaredMeanError += 2 * c * slopeCoefficients[j] * covarianceLowerTriangularMatrix[i][j];
         }
       }
     }
-    final double rss = squaredError;
+    final double mse = squaredMeanError;
     return new Statistics() {
       @Override
       public double getRss() {
-        return rss;
+        return mse * model.getStatsSampling().getCount();
       }
 
       @Override
       public double getMse() {
-        return rss / model.getStatsSampling().getCount();
+        return mse;
       }
 
       @Override
       public double getR2() {
-        return 1 - (rss / responseVariance);
+        return 1 - (getRss() / (responseVariance * model.getStatsSampling().getCount()));
       }
     };
   }

src/main/java/org/scaleborn/linereg/sampling/exact/ExactCoefficientSquareTermSampling.java

@@ -36,11 +36,11 @@ public class ExactCoefficientSquareTermSampling implements
    * instability as well as to arithmetic overflow.
    */
   private double[][] featuresProductSums;
-  private ExactSamplingContext context;
+  private final ExactSamplingContext context;
 
-  public ExactCoefficientSquareTermSampling(ExactSamplingContext context) {
+  public ExactCoefficientSquareTermSampling(final ExactSamplingContext context) {
     this.context = context;
-    int featuresCount = context.getFeaturesCount();
+    final int featuresCount = context.getFeaturesCount();
     this.featuresProductSums = new double[featuresCount][];
     for (int i = 0; i < featuresCount; i++) {
       this.featuresProductSums[i] = new double[featuresCount];
@@ -49,31 +49,31 @@ public ExactCoefficientSquareTermSampling(ExactSamplingContext context) {
 
   @Override
   public double[][] getCovarianceLowerTriangularMatrix() {
-    int featuresCount = this.context.getFeaturesCount();
-    long count = this.context.getCount();
-    double[][] covMatrix = new double[featuresCount][];
-    double[] averages = this.context.getFeaturesMean();
-    double[] featureSums = this.context.featureSums;
+    final int featuresCount = this.context.getFeaturesCount();
+    final long count = this.context.getCount();
+    final double[][] covMatrix = new double[featuresCount][];
+    final double[] averages = this.context.getFeaturesMean();
+    final double[] featureSums = this.context.featureSums;
     for (int i = 0; i < featuresCount; i++) {
-      double avgI = averages[i];
+      final double avgI = averages[i];
       covMatrix[i] = new double[featuresCount];
       // Iterate until "i" due to the covariance matrix is symmetric and
       // build only the lower triangle
       for (int j = 0; j <= i; j++) {
-        double avgJ = averages[j];
-        covMatrix[i][j] =
+        final double avgJ = averages[j];
+        covMatrix[i][j] = (
             this.featuresProductSums[i][j] - avgI * featureSums[j] - avgJ * featureSums[i]
-                + count * avgI * avgJ;
+                + count * avgI * avgJ) / count;
       }
     }
     return covMatrix;
   }
 
   @Override
   public void sample(final double[] featureValues, final double responseValue) {
-    int featuresCount = this.context.getFeaturesCount();
+    final int featuresCount = this.context.getFeaturesCount();
     for (int i = 0; i < featuresCount; i++) {
-      double vi = featureValues[i];
+      final double vi = featureValues[i];
       for (int j = 0; j < featuresCount; j++) {
         this.featuresProductSums[i][j] += vi * featureValues[j];
       }
@@ -82,7 +82,7 @@ public void sample(final double[] featureValues, final double responseValue) {
 
   @Override
   public void merge(final ExactCoefficientSquareTermSampling fromSample) {
-    int featuresCount = this.context.getFeaturesCount();
+    final int featuresCount = this.context.getFeaturesCount();
     for (int i = 0; i < featuresCount; i++) {
       for (int j = 0; j < featuresCount; j++) {
         this.featuresProductSums[i][j] += fromSample.featuresProductSums[i][j];

src/main/java/org/scaleborn/linereg/sampling/exact/ExactModelSamplingFactory.java

@@ -53,8 +53,9 @@ public ExactResponseVarianceTermSampling(
 
     @Override
     public double getResponseVariance() {
-      return this.context.responseSquareSum
-          - this.context.responseSum / this.context.getCount() * this.context.responseSum;
+      return (this.context.responseSquareSum
+          - this.context.responseSum / this.context.getCount() * this.context.responseSum)
+          / this.context.getCount();
     }
 
     @Override
@@ -106,10 +107,10 @@ public double[] getFeaturesResponseCovariance() {
       final double[] featuresMean = this.context.getFeaturesMean();
       final double responseMean = this.context.getResponseMean();
       for (int i = 0; i < featuresCount; i++) {
-        covariance[i] =
+        covariance[i] = (
             this.context.featuresResponseProductSum[i] - featuresMean[i] * this.context.responseSum
                 - responseMean * this.context.featureSums[i]
-                + count * featuresMean[i] * responseMean;
+                + count * featuresMean[i] * responseMean) / count;
       }
       return covariance;
     }