Mushroom Poisonous Prediction (Decision Tree) in CSharp

This example requires Emgu CV 2.0.0.0 which is available from SVN or Nightly build only

Emgu CV version >= 2.0 contains working DTree class (it fixes a bug in version 1.5.0.1's implementation). The following code is an unit test for the OpenCV DTree class, which is also a .Net port of the OpenCV mushroom.exe example.

using System.Drawing;
using Emgu.CV;
using Emgu.CV.Structure;
using Emgu.CV.ML;
using Emgu.CV.ML.Structure;

...

private void ReadMushroomData(out Matrix<float> data, out Matrix<float> response)
{
   string[] rows = System.IO.File.ReadAllLines("agaricus-lepiota.data");

   int varCount = rows[0].Split(',').Length - 1;
   data = new Matrix<float>(rows.Length, varCount);
   response = new Matrix<float>(rows.Length, 1);
   int count = 0;
   foreach (string row in rows)
   {
      string[] values = row.Split(',');
      Char c = System.Convert.ToChar(values[0]);
      response[count, 0] = System.Convert.ToInt32(c);
      for (int i = 1; i < values.Length; i++)
         data[count, i - 1] = System.Convert.ToByte(System.Convert.ToChar(values[i]));
      count++;
   }
}

[Test]
public void TestDTreesMushroom()
{
   Matrix<float> data, response;
   ReadMushroomData(out data, out response);

   //Use the first 80% of data as training sample
   int trainingSampleCount = (int)(data.Rows * 0.8);

   Matrix<Byte> varType = new Matrix<byte>(data.Cols + 1, 1);
   varType.SetValue((byte)MlEnum.VAR_TYPE.CATEGORICAL); //the data is categorical

   Matrix<byte> sampleIdx = new Matrix<byte>(data.Rows, 1);
   using (Matrix<byte> sampleRows = sampleIdx.GetRows(0, trainingSampleCount, 1))
      sampleRows.SetValue(255);

   float[] priors = new float[] {1, 0.5f};
   GCHandle priorsHandle = GCHandle.Alloc(priors, GCHandleType.Pinned);

   MCvDTreeParams param = new MCvDTreeParams();
   param.maxDepth = 8;
   param.minSampleCount = 10;
   param.regressionAccuracy = 0;
   param.useSurrogates = true;
   param.maxCategories = 15;
   param.cvFolds = 10;
   param.use1seRule = true;
   param.truncatePrunedTree = true;
   param.priors = priorsHandle.AddrOfPinnedObject();

   using (DTree dtree = new DTree())
   {
      bool success = dtree.Train(
         data,
         Emgu.CV.ML.MlEnum.DATA_LAYOUT_TYPE.ROW_SAMPLE,
         response,
         null,
         sampleIdx,
         varType,
         null,
         param);

      if (!success) return;
      double trainDataCorrectRatio = 0;
      double testDataCorrectRatio = 0;
      for (int i = 0; i < data.Rows; i++)
      {
         using (Matrix<float> sample = data.GetRow(i))
         {
            double r = dtree.Predict(sample, null, false).value;
            r = Math.Abs(r - response[i, 0]);
            if (r < 1.0e-5)
            {
               if (i < trainingSampleCount)
                  trainDataCorrectRatio++;
               else
                  testDataCorrectRatio++;
            }
         }
      }

      trainDataCorrectRatio /= trainingSampleCount;
      testDataCorrectRatio /= (data.Rows - trainingSampleCount);

      Trace.WriteLine(String.Format("Prediction accuracy for training data :{0}%", trainDataCorrectRatio*100));
      Trace.WriteLine(String.Format("Prediction accuracy for test data :{0}%", testDataCorrectRatio*100));
   }

   priorsHandle.Free();
}

The result of running this unit test:

Prediction accuracy for training data :99.8769041390983%
Prediction accuracy for test data :99.2615384615385%

That's a really good prediction rate. A big thanks to OpenCV developers.