public void Learn()
{
var rootMask = new BitArrayX(input.Length, true);
Root = new DecisionNode
{
SubsetMask = new SubsetMask(rootMask, input.Length, output, outputFactorsCount),
UsedFactors = new BitArrayX(inputFactorsCount)
};
var stack = new Stack <DecisionNode>();
stack.Push(Root);
while (stack.Count > 0)
{
var node = stack.Pop();
SplitNode(node);
if (node.IsLeaf)
{
continue;
}
stack.Push(node.Present);
stack.Push(node.Absent);
}
}
private static IEnumerable <DecisionNode> GetChildren(DecisionNode node)
{
if (!node.IsLeaf)
{
yield return(node.Present);
yield return(node.Absent);
}
}
private string ToString(DecisionNode node)
{
var factor = node.IsLeaf
? OutputFactors[node.Factor]
: InputFactors[node.Factor];
return($"[{indices[node]}] " +
(node.IsLeaf
? $"Return {factor}"
: $"If {factor} present {indices[node.Present]}, absent {indices[node.Absent]}"));
}
private void SplitNode(DecisionNode node)
{
var mask = node.SubsetMask.Mask;
var size = node.SubsetMask.Size;
if (node.SubsetMask.EntropyZeroFactor >= 0)
{
node.IsLeaf = true;
node.Factor = node.SubsetMask.EntropyZeroFactor;
return;
}
var bestInformationGain = double.MinValue;
var bestFactor = -1;
var bestPresentSplitMask = default(SubsetMask);
var bestAbsentSplitMask = default(SubsetMask);
for (var inputFactor = 0; inputFactor < inputFactorsCount; inputFactor++)
{
if (node.UsedFactors[inputFactor])
{
continue;
}
var factorPresentMask = new BitArrayX(mask).And(factorPresentMasks[inputFactor]);
var factorAbsentMask = new BitArrayX(mask).AndNot(factorPresentMask);
var factorPresentCount = factorPresentMask.CountBitSet();
var factorAbsentCount = size - factorPresentCount;
var factorPresentSubset = new SubsetMask(factorPresentMask, factorPresentCount, output, outputFactorsCount);
var factorAbsentSubset = new SubsetMask(factorAbsentMask, factorAbsentCount, output, outputFactorsCount);
var informationGain = node.SubsetMask.Entropy -
(factorPresentSubset.Entropy * factorPresentCount / size) -
(factorAbsentSubset.Entropy * factorAbsentCount / size);
if (informationGain > bestInformationGain)
{
bestInformationGain = informationGain;
bestFactor = inputFactor;
bestPresentSplitMask = factorPresentSubset;
bestAbsentSplitMask = factorAbsentSubset;
}
}
if (bestFactor == -1)
{
var counts = new int[outputFactorsCount];
for (var i = 0; i < output.Length; i++)
{
if (mask[i])
{
counts[output[i]]++;
}
}
node.IsLeaf = true;
node.Factor = counts.IndexOfMax();
return;
}
var childUsedFactors = new BitArrayX(node.UsedFactors)
{
[bestFactor] = true
};
node.Factor = bestFactor;
node.Present = new DecisionNode
{
SubsetMask = bestPresentSplitMask,
Factor = bestFactor,
Parent = node,
UsedFactors = childUsedFactors,
};
node.Absent = new DecisionNode
{
SubsetMask = bestAbsentSplitMask,
Factor = bestFactor,
Parent = node,
UsedFactors = childUsedFactors,
};
}
