/// <summary>
/// Compute sample weights such that the class distribution of y becomes
/// balanced.
/// </summary>
/// <param name="?"></param>
private static double[] BalanceWeights(int[] y)
{
var encoder = new LabelEncoder <int>();
y = encoder.FitTransform(y);
var bins = Np.BinCount(y);
var weights = bins.ElementsAt(y).Select(v => 1.0 / v * bins.Min()).ToArray();
return(weights);
}
public void TestMinSamplesLeaf()
{
foreach (var name in CLF_TREES)
{
var est = CreateClassifier <double>(name, min_samples_leaf: 5, random: new Random(0));
est.Fit(X, y);
var @out = est.Tree.apply(X.ToDenseMatrix());
var node_counts = Np.BinCount(@out.Select(v => (int)v).ToArray());
var leaf_count = node_counts.Where(v => v != 0).ToList(); // drop inner nodes
Assert.IsTrue(leaf_count.Min() > 4,
"Failed with {0}".Frmt(name));
}
foreach (var name in RegTrees)
{
var est = CreateRegressor(name, min_samples_leaf: 5, random: new Random(0));
est.Fit(X, y);
var @out = est.Tree.apply(X.ToDenseMatrix());
var nodeCounts = Np.BinCount(@out.Select(v => (int)v).ToArray());
var leafCount = nodeCounts.Where(v => v != 0).ToList(); // drop inner nodes
Assert.IsTrue(leafCount.Min() > 4,
"Failed with {0}".Frmt(name));
}
}
public void TestSampleWeight()
{
// Test that zero-weighted samples are not taken into account
var X = Enumerable.Range(0, 100).ToColumnMatrix();
var y = Enumerable.Repeat(1, 100).ToArray();
Array.Clear(y, 0, 50);
var sampleWeight = Enumerable.Repeat(1, 100).ToVector();
sampleWeight.SetSubVector(0, 50, Enumerable.Repeat(0, 50).ToVector());
var clf = new DecisionTreeClassifier <int>(random: new Random(0));
clf.Fit(X, y, sampleWeight: sampleWeight);
AssertExt.ArrayEqual(clf.Predict(X), Enumerable.Repeat(1, 100).ToArray());
// Test that low weighted samples are not taken into account at low depth
X = Enumerable.Range(0, 200).ToColumnMatrix();
y = new int[200];
Array.Copy(Enumerable.Repeat(1, 50).ToArray(), 0, y, 50, 50);
Array.Copy(Enumerable.Repeat(2, 100).ToArray(), 0, y, 100, 100);
X.SetSubMatrix(100, 100, 0, 1, Enumerable.Repeat(200, 100).ToColumnMatrix());
sampleWeight = Enumerable.Repeat(1, 200).ToVector();
sampleWeight.SetSubVector(100, 100, Enumerable.Repeat(0.51, 100).ToVector());
// Samples of class '2' are still weightier
clf = new DecisionTreeClassifier <int>(maxDepth: 1, random: new Random(0));
clf.Fit(X, y, sampleWeight: sampleWeight);
Assert.AreEqual(149.5, clf.Tree.Threshold[0]);
sampleWeight.SetSubVector(100, 100, Enumerable.Repeat(0.50, 100).ToVector());
// Samples of class '2' are no longer weightier
clf = new DecisionTreeClassifier <int>(maxDepth: 1, random: new Random(0));
clf.Fit(X, y, sampleWeight: sampleWeight);
Assert.AreEqual(49.5, clf.Tree.Threshold[0]); // Threshold should have moved
// Test that sample weighting is the same as having duplicates
X = iris.Data;
y = iris.Target;
var random = new Random(0);
var duplicates = new int[200];
for (int i = 0; i < duplicates.Length; i++)
{
duplicates[i] = random.Next(X.RowCount);
}
clf = new DecisionTreeClassifier <int>(random: new Random(1));
clf.Fit(X.RowsAt(duplicates), y.ElementsAt(duplicates));
sampleWeight = Np.BinCount(duplicates, minLength: X.RowCount).ToVector();
var clf2 = new DecisionTreeClassifier <int>(random: new Random(1));
clf2.Fit(X, y, sampleWeight: sampleWeight);
var @internal = clf.Tree.ChildrenLeft.Indices(v => v != Tree._TREE_LEAF);
AssertExt.AlmostEqual(clf.Tree.Threshold.ElementsAt(@internal),
clf2.Tree.Threshold.ElementsAt(@internal));
}