private void ProcessData(
string dependentFeatureName,
Matrix <double> trainingData,
IList <TPredictionResult> expectedValues,
IList <string> dataColumnNames,
ConcurrentDictionary <int, TPredictionResult> partialResults,
IKnnAdditionalParams knnAdditionalParams)
{
Parallel.For(0, trainingData.RowCount, rowIdx =>
{
var trainingDataExceptRow = trainingData.RemoveRow(rowIdx);
var expectedValuesExceptRow = expectedValues.ToList();
expectedValuesExceptRow.RemoveAt(rowIdx);
var queryMatrix = Matrix <double> .Build.DenseOfRowVectors(trainingData.Row(rowIdx));
var queryDataFrame = new DataFrame(queryMatrix);
var knnPredictionModel = new KnnPredictionModel <TPredictionResult>(trainingDataExceptRow, expectedValuesExceptRow,
dataColumnNames, knnAdditionalParams.KNeighbors, knnAdditionalParams.UseWeightedDistances);
var results = _knnPredictor.Predict(
queryDataFrame,
knnPredictionModel,
dependentFeatureName);
TPredictionResult result = results.First();
partialResults.AddOrUpdate(rowIdx, result, (i, d) => result);
});
}
private double ProcessDataAndQuantifyErrorRate(
string dependentFeatureName,
Matrix <double> trainingData,
IList <TPredictionResult> expectedValues,
IList <string> dataColumnsNames,
IKnnAdditionalParams knnAdditionalParams)
{
var partialResults = new ConcurrentDictionary <int, TPredictionResult>();
ProcessData(dependentFeatureName, trainingData, expectedValues, dataColumnsNames, partialResults, knnAdditionalParams);
var actualValues = partialResults.OrderBy(kvp => kvp.Key).Select(kvp => kvp.Value).ToList();
return(_errorMeasure.CalculateError(expectedValues, actualValues));
}
protected IBackwardsEliminationKnnModel <TPredictionResult> PerformBackwardsElimination(IDataFrame dataFrame, string dependentFeatureName, IKnnAdditionalParams additionalParams)
{
Tuple <Matrix <double>, IList <TPredictionResult>, IList <string> > preparedData = PrepareTrainingData(dataFrame, dependentFeatureName);
var dataColumnsNames = preparedData.Item3;
var trainingData = _dataNormalizer.NormalizeColumns(preparedData.Item1);
var expectedValues = preparedData.Item2;
// TODO: refactor this - ugly!
_knnPredictor.NormalizeNumericValues = false;
double baseErrorRate = ProcessDataAndQuantifyErrorRate(
dependentFeatureName,
trainingData,
expectedValues,
dataColumnsNames,
additionalParams);
var actualDataColumnNames = new List <string>(dataColumnsNames);
var anyFeatureRemovedInThisIteration = true;
var removedFeaturesInfo = new List <IBackwardsEliminationRemovedFeatureData>();
while (anyFeatureRemovedInThisIteration)
{
anyFeatureRemovedInThisIteration = false;
var candidateFeaturesToEliminate = new Dictionary <int, double>();
foreach (var columnIdx in Enumerable.Range(0, actualDataColumnNames.Count))
{
var newFeatureNames = new List <string>(actualDataColumnNames);
newFeatureNames.RemoveAt(columnIdx);
var trainingDataWithoutColumn = trainingData.RemoveColumn(columnIdx);
var newDataPredictionError = ProcessDataAndQuantifyErrorRate(
dependentFeatureName,
trainingDataWithoutColumn,
expectedValues,
newFeatureNames,
additionalParams);
if (newDataPredictionError <= baseErrorRate)
{
var errorGain = baseErrorRate - newDataPredictionError;
candidateFeaturesToEliminate.Add(columnIdx, errorGain);
}
}
if (!candidateFeaturesToEliminate.Any())
{
break;
}
var bestFeatureToRemove = candidateFeaturesToEliminate.OrderBy(kvp => kvp.Value).First();
anyFeatureRemovedInThisIteration = true;
removedFeaturesInfo.Add(new BackwardsEliminationRemovedFeatureData(bestFeatureToRemove.Value, actualDataColumnNames[bestFeatureToRemove.Key]));
actualDataColumnNames.RemoveAt(bestFeatureToRemove.Key);
baseErrorRate = bestFeatureToRemove.Value;
}
return(new BackwardsEliminationKnnModel <TPredictionResult>(
preparedData.Item1,
expectedValues,
dataColumnsNames,
additionalParams.KNeighbors,
additionalParams.UseWeightedDistances,
removedFeaturesInfo));
}
