private static double GetGainForAllAttributes(TrainingData data, int colIndex, double entropyOfDataset)
{
var totalRows = data.Rows.Length;
var amountForDifferentValue = GetAmountOfEdgesAndTotalPositivResults(data, colIndex);
var stepsForCalculation = new List <double>();
foreach (var item in amountForDifferentValue)
{
// helper for calculation
var firstDivision = item[0, 1] / (double)item[0, 0];
var secondDivision = (item[0, 0] - item[0, 1]) / (double)item[0, 0];
// prevent dividedByZeroException
if (firstDivision == 0 || secondDivision == 0)
{
stepsForCalculation.Add(0.0);
}
else
{
stepsForCalculation.Add(-firstDivision * Math.Log(firstDivision, 2) - secondDivision * Math.Log(secondDivision, 2));
}
}
var gain = stepsForCalculation.Select((t, i) => amountForDifferentValue[i][0, 0] / (double)totalRows * t).Sum();
gain = entropyOfDataset - gain;
return(gain);
}
private static List <int[, ]> GetAmountOfEdgesAndTotalPositivResults(TrainingData data, int indexOfColumnToCheck)
{
var foundValues = new List <int[, ]>();
var knownValues = data.Rows.Select(x => x[indexOfColumnToCheck]).Distinct().ToList();
foreach (var item in knownValues)
{
var amount = 0;
var positiveAmount = 0;
for (var i = 0; i < data.Rows.Length; i++)
{
if (data.Rows[i][indexOfColumnToCheck].ToString().Equals(item))
{
amount++;
// Counts the positive cases and adds the sum later to the array for the calculation
if (data.Rows[i][data.Columns.Length - 1].ToString().Equals(data.Rows[0][data.Columns.Length - 1]))
{
positiveAmount++;
}
}
}
int[,] array = { { amount, positiveAmount } };
foundValues.Add(array);
}
return(foundValues);
}
private static TreeNode GetRootNode(TrainingData data, string edge)
{
var attributes = new List <NodeAttribute>();
var highestInformationGainIndex = -1;
var highestInformationGain = double.MinValue;
// Get all names, amount of attributes and attributes for every column
for (var i = 0; i < data.Columns.Length - 1; i++)
{
var differentAttributenames = data.Rows.Select(x => x[i]).Distinct().ToList();
attributes.Add(new NodeAttribute(data.Columns[i].ToString(), differentAttributenames));
}
// Calculate Entropy (S)
var tableEntropy = CalculateEntropy(data);
for (var i = 0; i < attributes.Count; i++)
{
attributes[i].InformationGain = GetGainForAllAttributes(data, i, tableEntropy);
if (attributes[i].InformationGain > highestInformationGain)
{
highestInformationGain = attributes[i].InformationGain;
highestInformationGainIndex = i;
}
}
return(new TreeNode(attributes[highestInformationGainIndex].Name, highestInformationGainIndex, attributes[highestInformationGainIndex], edge));
}
private static bool CheckIfIsLeaf(TreeNode root, TrainingData data, string attributeToCheck)
{
var isLeaf = true;
var allEndValues = new List <string>();
// get all leaf values for the attribute in question
for (var i = 0; i < data.Rows.Length; i++)
{
if (data.Rows[i][root.TableIndex].ToString().Equals(attributeToCheck))
{
allEndValues.Add(data.Rows[i][data.Columns.Length - 1].ToString());
}
}
// check whether all elements of the list have the same value
if (allEndValues.Count > 0 && allEndValues.Any(x => x != allEndValues[0]))
{
isLeaf = false;
}
// create leaf with value to display and edge to the leaf
if (isLeaf)
{
root.ChildNodes.Add(new TreeNode(true, allEndValues[0], attributeToCheck));
}
return(isLeaf);
}
public static TreeNode Learn(TrainingData data, string edgeName)
{
var root = GetRootNode(data, edgeName);
foreach (var item in root.NodeAttribute.DifferentAttributeNames)
{
// if a leaf, leaf will be added in this method
var isLeaf = CheckIfIsLeaf(root, data, item);
// make a recursive call as long as the node is not a leaf
if (!isLeaf)
{
var reducedTable = new TrainingData
{
Columns = data.Columns.Skip(root.TableIndex + 1).Select(x => x).ToArray(),
Rows = data.Rows.Where(x => x[root.TableIndex].Equals(item))
.Select(x => x.Skip(root.TableIndex + 1).ToArray())
.ToArray()
};
root.ChildNodes.Add(Learn(reducedTable, item));
}
}
return(root);
}
private static double CalculateEntropy(TrainingData data)
{
var totalRows = data.Rows.Length;
var amountForDifferentValue = GetAmountOfEdgesAndTotalPositivResults(data, data.Columns.Length - 1);
var stepsForCalculation = amountForDifferentValue
.Select(item => item[0, 0] / (double)totalRows)
.Select(division => - division * Math.Log(division, 2))
.ToList();
return(stepsForCalculation.Sum());
}
public TrainingData Read(string filePath)
{
var data = new TrainingData {
};
string[] lines = File.ReadAllLines(filePath);
data.Rows = new string[lines.Length - 1][];
for (int rowIndex = 0; rowIndex < lines.Length; rowIndex++)
{
string line = lines[rowIndex];
// Columns
if (rowIndex == 0)
{
data.Columns = line.Split(';').Where(x => !String.IsNullOrEmpty(x)).ToArray();
}
else
{
data.Rows[rowIndex - 1] = line.Split(';').Where(x => !String.IsNullOrEmpty(x)).ToArray();
}
}
return(data);
}