Changeset 2136 for trunk/sources/HeuristicLab.Modeling

Timestamp:

07/06/09 17:12:37 (15 years ago)

Author:

gkronber

Message:

Improved handling of exceptional cases in data-based modeling evaluators. #688 (SimpleEvaluators should handle exceptional cases more gracefully)

Location:

trunk/sources/HeuristicLab.Modeling/3.2

Files:

• SimpleEvaluatorBase.cs (modified) (1 diff)
• SimpleMSEEvaluator.cs (modified) (2 diffs)
• SimpleMeanAbsolutePercentageErrorEvaluator.cs (modified) (2 diffs)
• SimpleMeanAbsolutePercentageOfRangeErrorEvaluator.cs (modified) (3 diffs)
• SimpleR2Evaluator.cs (modified) (2 diffs)
• SimpleVarianceAccountedForEvaluator.cs (modified) (2 diffs)
• VariableEvaluationImpactCalculator.cs (modified) (3 diffs)

trunk/sources/HeuristicLab.Modeling/3.2/SimpleEvaluatorBase.cs

@@ -31 +31 @@
 
     public abstract double Evaluate(double[,] values);
+
+    protected static bool IsAlmost(double x, double y) {
+      return Math.Abs(x - y) < 1.0E-12;
+    }
   }
 }

trunk/sources/HeuristicLab.Modeling/3.2/SimpleMSEEvaluator.cs

@@ -16 +16 @@
     }
     public override double Evaluate(double[,] values) {
-      return Calculate(values);
+      try {
+        return Calculate(values);
+      }
+      catch (ArgumentException) {
+        return double.PositiveInfinity;
+      }
     }
 
@@ -32 +37 @@
         }
       }
-
-      double mse = sse / cnt;
-      return mse;
+      if (cnt > 0) {
+        double mse = sse / cnt;
+        return mse;
+      } else {
+        throw new ArgumentException("Mean squared errors is not defined for input vectors of NaN or Inf");
+      }
     }
   }

trunk/sources/HeuristicLab.Modeling/3.2/SimpleMeanAbsolutePercentageErrorEvaluator.cs

@@ -38 +38 @@
 
     public override double Evaluate(double[,] values) {
-      return Calculate(values);
+      try {
+        return Calculate(values);
+      }
+      catch (ArgumentException) {
+        return double.PositiveInfinity;
+      }
     }
 
@@ -55 +60 @@
         }
       }
-      return errorsSum / n;
+      if (n > 0) {
+        return errorsSum / n;
+      } else throw new ArgumentException("Mean of absolute percentage error is not defined for input vectors of NaN or Inf");
     }
   }

trunk/sources/HeuristicLab.Modeling/3.2/SimpleMeanAbsolutePercentageOfRangeErrorEvaluator.cs

@@ -37 +37 @@
 
     public override double Evaluate(double[,] values) {
-      return Calculate(values);
+      try {
+        return Calculate(values);
+      }
+      catch (ArgumentException) {
+        return double.PositiveInfinity;
+      }
     }
 
@@ -47 +52 @@
       for (int i = 0; i < originalValues.Length; i++) originalValues[i] = values[i, 1];
       double range = Statistics.Range(originalValues);
+      if (double.IsInfinity(range)) throw new ArgumentException("Range of elements in values is infinity");
+      if (IsAlmost(range, 0.0)) throw new ArgumentException("Range of elements in values is zero");
 
       for (int i = 0; i < values.GetLength(0); i++) {
@@ -59 +66 @@
         }
       }
-      return errorsSum / n;
+      if (double.IsInfinity(range) || n == 0) {
+        throw new ArgumentException("Mean of absolute percentage of range error is not defined for input vectors of NaN or Inf");
+      } else {
+        return errorsSum / n;
+      }
     }
   }

trunk/sources/HeuristicLab.Modeling/3.2/SimpleR2Evaluator.cs

@@ -17 +17 @@
 
     public override double Evaluate(double[,] values) {
-      return Calculate(values);
+      try {
+        return Calculate(values);
+      }
+      catch (ArgumentException) {
+        return double.NegativeInfinity;
+      }
     }
 
@@ -35 +40 @@
         }
       }
-      targetMean /= cnt;
 
-      double targetDeviationTotalSumOfSquares = 0;
-      for (int i = 0; i < values.GetLength(0); i++) {
-        double target = values[i, 1];
-        if (!double.IsNaN(target) && !double.IsInfinity(target)) {
-          target = target - targetMean;
-          target = target * target;
-          targetDeviationTotalSumOfSquares += target;
+      if (cnt > 0) {
+        targetMean /= cnt;
+
+        double targetDeviationTotalSumOfSquares = 0;
+        for (int i = 0; i < values.GetLength(0); i++) {
+          double target = values[i, 1];
+          if (!double.IsNaN(target) && !double.IsInfinity(target)) {
+            target = target - targetMean;
+            target = target * target;
+            targetDeviationTotalSumOfSquares += target;
+          }
         }
+        double quality = 1 - sse / targetDeviationTotalSumOfSquares;
+        if (quality > 1)
+          throw new InvalidProgramException();
+
+        return quality;
+      } else {
+        throw new ArgumentException("Coefficient of determination is not defined for input vectors of NaN or Inf");
       }
-      double quality = 1 - sse / targetDeviationTotalSumOfSquares;
-      if (quality > 1)
-        throw new InvalidProgramException();
-
-      return quality;
     }
   }

trunk/sources/HeuristicLab.Modeling/3.2/SimpleVarianceAccountedForEvaluator.cs

@@ -43 +43 @@
 
     public override double Evaluate(double[,] values) {
-      return Calculate(values);
+      try {
+        return Calculate(values);
+      }
+      catch (ArgumentException) {
+        return double.NegativeInfinity;
+      }
     }
 
@@ -64 +69 @@
       double errorsVariance = Statistics.Variance(errors);
       double originalsVariance = Statistics.Variance(originalTargetVariableValues);
-      double quality = 1 - errorsVariance / originalsVariance;
-
-      return quality;
+      if (IsAlmost(originalsVariance, 0.0))
+        if (IsAlmost(errorsVariance, 0.0)) {
+          return 1.0;
+        } else {
+          throw new ArgumentException("Variance of original values is zero");
+        } else {
+        return 1.0 - errorsVariance / originalsVariance;
+      }
     }
   }

trunk/sources/HeuristicLab.Modeling/3.2/VariableEvaluationImpactCalculator.cs

@@ -50 +50 @@
 
     protected override double CalculateImpact(double[] referenceValue, double[] newValue) {
-      return SimpleMSEEvaluator.Calculate(CombineOutputs(referenceValue, newValue));
+      try {
+        return SimpleMSEEvaluator.Calculate(CombineOutputs(referenceValue, newValue));
+      }
+      catch (ArgumentException) {
+        return double.PositiveInfinity;
+      }
     }
 
@@ -59 +64 @@
     protected override double[] PostProcessImpacts(double[] impacts) {
       double mseSum = impacts.Sum();
-      if (mseSum == 0.0) mseSum = 1.0;
+      if (IsAlmost(mseSum, 0.0)) mseSum = 1.0;
       for (int i = 0; i < impacts.Length; i++) {
         impacts[i] = impacts[i] / mseSum;
@@ -66 +71 @@
     }
 
+    private bool IsAlmost(double x, double y) {
+      return Math.Abs(x - y) < 1.0E-12;
+    }
+
     protected abstract double[] GetOutputs(IScope scope, Dataset dataset, int targetVariable, ItemList<IntData> allowedFeatures, int start, int end);
   }