private static DecisionTreeNode <T, D> computeDepthFirst(DecisionTreeNode <T, D> node, List <T> data, IFeatureFactory <T, D> factory, int numFeatures, int numThresholds, int numLabels, float[] labelWeights, int depth)
{
GC.Collect();
if (data.Count == 0)
{
UpdateManager.WriteLine("No data at depth {0}", depth);
return(null);
}
if (data[0] is IComparable <T> )
{
data.Sort();
}
if (checkDelta(data))
{
UpdateManager.WriteLine("Delta function at depth {0}", depth);
int label = data[0].Label;
float[] dist = new float[numLabels];
dist[label] = 1;
return(new DecisionTreeNode <T, D>(dist));
}
int dataCount = data.Count;
Decider <T, D> bestDecider = null;
float bestScore = float.MinValue;
float[] bestLeftDistribution = null;
float[] bestRightDistribution = null;
using (ThreadLocal <DecisionResult> results = new ThreadLocal <DecisionResult>(() => new DecisionResult {
Score = bestScore
}, true))
{
Parallel.For(0, numFeatures, i =>
{
float[] leftDistribution;
float[] rightDistribution;
Decider <T, D> decider = new Decider <T, D>(factory);
decider.LoadData(data);
float score = decider.ChooseThreshold(numThresholds, numLabels, labelWeights, out leftDistribution, out rightDistribution);
if (score > results.Value.Score)
{
results.Value = new DecisionResult {
LeftDistribution = leftDistribution, RightDistribution = rightDistribution, Decider = decider, Score = score
};
}
});
foreach (var result in results.Values)
{
if (result.Score > bestScore)
{
bestLeftDistribution = result.LeftDistribution;
bestRightDistribution = result.RightDistribution;
bestDecider = result.Decider;
bestScore = result.Score;
}
}
}
float support = 0;
if (labelWeights != null)
{
foreach (T point in data)
{
support += labelWeights[point.Label];
}
}
else
{
support = dataCount;
}
if (bestScore == float.MinValue || dataCount < MinimumSupport)
{
UpdateManager.WriteLine("Stopping due to lack of data at depth {0}, {1} < {2}", depth, dataCount, MinimumSupport);
float[] distribution = new float[numLabels];
for (int i = 0; i < dataCount; i++)
{
distribution[data[i].Label]++;
}
if (labelWeights != null)
{
for (int i = 0; i < distribution.Length; i++)
{
distribution[i] *= labelWeights[i];
}
}
return(new DecisionTreeNode <T, D>(distribution));
}
if (depth == MaximumDepth - 2)
{
UpdateManager.WriteLine("Last branch node trained at depth {0}", depth);
node.Left = new DecisionTreeNode <T, D>(bestLeftDistribution);
node.Right = new DecisionTreeNode <T, D>(bestRightDistribution);
node.NodeType = NodeType.Branch;
node.Decider = bestDecider;
return(node);
}
Decision[] decisions = bestDecider.Decide(data);
List <T> leftData = new List <T>();
List <T> rightData = new List <T>();
for (int i = 0; i < decisions.Length; i++)
{
if (decisions[i] == Decision.Left)
{
leftData.Add(data[i]);
}
else
{
rightData.Add(data[i]);
}
}
if (leftData.Count == 0 || rightData.Count == 0)
{
throw new Exception("Error");
}
UpdateManager.WriteLine("Branch node at depth {0} trained.", depth);
node.Left = computeDepthFirst(new DecisionTreeNode <T, D>(), leftData, factory, numFeatures, numThresholds, numLabels, labelWeights, depth + 1);
node.Right = computeDepthFirst(new DecisionTreeNode <T, D>(), rightData, factory, numFeatures, numThresholds, numLabels, labelWeights, depth + 1);
node.Decider = bestDecider;
node.NodeType = NodeType.Branch;
return(node);
}
private static DecisionTreeNode <T, D> computeBreadthFirst(float threshold, List <T> data, IFeatureFactory <T, D> factory, int numFeatures, int numThresholds, int numLabels, float[] labelWeights)
{
string id = "DecisionTree.ComputeBreadthFirst";
Queue <SplitCandidate> candidates = new Queue <SplitCandidate>();
SplitCandidate start = new SplitCandidate(new List <int>(), 1, 0);
for (int i = 0; i < data.Count; i++)
{
start.Indices.Add(i);
}
start.Entropy = calculateEntropy(data, start.Indices, labelWeights, numLabels);
start.Support = calculateSupport(data, start.Indices, labelWeights);
candidates.Enqueue(start);
bool changed = true;
float[] leftDistribution, rightDistribution;
int tries = (int)_numberOfTries;
float increment = threshold / tries;
Dictionary <int, Decider <T, D> > deciders = new Dictionary <int, Decider <T, D> >();
while (tries > 0)
{
if (!changed)
{
threshold -= increment;
UpdateManager.WriteLine("Decreasing threshold to {0}", threshold);
}
GC.Collect();
int count = candidates.Count;
for (int i = 0; i < count; i++)
{
SplitCandidate candidate = candidates.Dequeue();
if (candidate.Delta)
{
candidates.Enqueue(candidate);
continue;
}
if (MaximumDepth > 0 && candidate.Level >= MaximumDepth - 1)
{
candidates.Enqueue(candidate);
continue;
}
if (candidate.Support < MinimumSupport)
{
candidates.Enqueue(candidate);
continue;
}
int dataCount = candidate.Indices.Count;
if (candidate.Values == null)
{
candidate.Values = new float[dataCount];
}
if (candidate.Labels == null)
{
candidate.Labels = new int[dataCount];
}
if (candidate.Weights == null)
{
candidate.Weights = new float[dataCount];
}
candidates.Enqueue(candidate);
}
float bestGain = float.MinValue;
for (int k = 0; k < numFeatures; k++)
{
UpdateManager.RaiseProgress(k, numFeatures);
Decider <T, D> decider = new Decider <T, D>(factory);
decider.ApplyFeature(data);
for (int i = 0; i < count; i++)
{
SplitCandidate candidate = candidates.Dequeue();
if (MaximumDepth > 0 && candidate.Level >= MaximumDepth - 1)
{
candidates.Enqueue(candidate);
continue;
}
if (candidate.Delta)
{
candidates.Enqueue(candidate);
continue;
}
if (candidate.Support < MinimumSupport)
{
candidates.Enqueue(candidate);
continue;
}
List <int> indices = candidate.Indices;
int dataCount = indices.Count;
for (int j = 0; j < dataCount; j++)
{
T point = data[indices[j]];
candidate.Values[j] = point.FeatureValue;
candidate.Labels[j] = point.Label;
candidate.Weights[j] = point.Weight;
}
decider.SetData(candidate.Values, candidate.Weights, candidate.Labels);
float gain = candidate.Entropy + decider.ChooseThreshold(numThresholds, numLabels, labelWeights, out leftDistribution, out rightDistribution);
bestGain = Math.Max(gain, bestGain);
if ((gain > threshold || candidate.Level < MinimumDepth) && gain > candidate.EntropyGain)
{
candidate.EntropyGain = gain;
candidate.Decider = new Decider <T, D>(decider.Feature, decider.Threshold);
}
candidates.Enqueue(candidate);
}
}
UpdateManager.WriteLine(id, "\rNodes Added:");
changed = false;
for (int i = 0; i < count; i++)
{
SplitCandidate candidate = candidates.Dequeue();
if (candidate.Decider == null)
{
candidates.Enqueue(candidate);
continue;
}
changed = true;
List <int> indices = candidate.Indices;
int dataCount = candidate.Indices.Count;
List <T> points = new List <T>();
for (int j = 0; j < dataCount; j++)
{
points.Add(data[indices[j]]);
}
Decision[] decisions = candidate.Decider.Decide(points);
List <int> left = new List <int>();
List <int> right = new List <int>();
for (int j = 0; j < dataCount; j++)
{
if (decisions[j] == Decision.Left)
{
left.Add(indices[j]);
}
else
{
right.Add(indices[j]);
}
}
SplitCandidate leftCandidate = new SplitCandidate(left, 2 * candidate.Index, candidate.Level + 1);
SplitCandidate rightCandidate = new SplitCandidate(right, 2 * candidate.Index + 1, candidate.Level + 1);
leftCandidate.Entropy = calculateEntropy(data, left, labelWeights, numLabels);
leftCandidate.Support = calculateSupport(data, left, labelWeights);
leftCandidate.Delta = calculateDelta(data, left);
rightCandidate.Entropy = calculateEntropy(data, right, labelWeights, numLabels);
rightCandidate.Support = calculateSupport(data, right, labelWeights);
rightCandidate.Delta = calculateDelta(data, right);
UpdateManager.WriteLine(id, "{3:00000}:{0:0.000}|{1:0.000} {2:0.000} {4}", leftCandidate.Support / candidate.Support, rightCandidate.Support / candidate.Support, candidate.EntropyGain, candidate.Index, candidate.Decider);
deciders[candidate.Index] = candidate.Decider;
candidates.Enqueue(leftCandidate);
candidates.Enqueue(rightCandidate);
}
if (!changed)
{
UpdateManager.WriteLine("No new nodes added, best entropy gain was {0}", bestGain);
tries--;
}
if (bestGain == float.MinValue)
{
break;
}
}
Dictionary <int, List <int> > leafIndices = new Dictionary <int, List <int> >();
while (candidates.Count > 0)
{
SplitCandidate candidate = candidates.Dequeue();
leafIndices[candidate.Index] = candidate.Indices;
}
return(buildTree(new DecisionTreeNode <T, D>(), 1, deciders, leafIndices, data, numLabels, labelWeights));
}