private Measure[] CloneMeasure(Measure[] measures)
{
var m = new Measure[measures.Length];
for (int i = 0; i < m.Length; i++)
m[i] = measures[i].Clone();
return m;
}
/// <summary>Generates an object from its XML representation.</summary>
/// <param name="reader">The <see cref="T:System.Xml.XmlReader" /> stream from which the object is
/// deserialized.</param>
public override void ReadXml(XmlReader reader)
{
reader.MoveToContent();
reader.ReadStartElement();
Descriptor = Xml.Read<Descriptor>(reader);
Root = Xml.Read<Measure>(reader);
}
/// <summary>Generate model based on a set of examples.</summary>
/// <exception cref="InvalidOperationException">Thrown when the requested operation is invalid.</exception>
/// <param name="X">The Matrix to process.</param>
/// <param name="y">The Vector to process.</param>
/// <returns>Model.</returns>
public override IModel Generate(Matrix X, Vector y)
{
if (Descriptor == null)
throw new InvalidOperationException("Cannot build naive bayes model without type knowledge!");
// create answer probabilities
if (!Descriptor.Label.Discrete)
throw new InvalidOperationException("Need to use regression for non-discrete labels!");
this.Preprocess(X);
// compute Y probabilities
Statistic[] statistics = GetLabelStats(y);
Measure root = new Measure
{
Discrete = true,
Label = Descriptor.Label.Name,
Probabilities = statistics
};
// collect feature ranges
Measure[] features = GetBaseConditionals(X);
// compute conditional counts
for (int i = 0; i < y.Length; i++)
{
var stat = statistics.Where(s => s.X.Min == y[i]).First();
if (stat.Conditionals == null)
stat.Conditionals = CloneMeasure(features);
for (int j = 0; j < X.Cols; j++)
{
var s = stat.Conditionals[j];
s.Increment(X[i, j]);
}
}
// normalize into probabilities
for (int i = 0; i < statistics.Length; i++)
{
var cond = statistics[i];
for (int j = 0; j < cond.Conditionals.Length; j++)
cond.Conditionals[j].Normalize();
}
// label ids
LabelIds(root);
return new NaiveBayesModel
{
Descriptor = Descriptor,
NormalizeFeatures = base.NormalizeFeatures,
FeatureNormalizer = base.FeatureNormalizer,
FeatureProperties = base.FeatureProperties,
Root = root
};
}
/// <summary>Generate model based on a set of examples.</summary>
/// <exception cref="InvalidOperationException">Thrown when the requested operation is invalid.</exception>
/// <param name="x">The Matrix to process.</param>
/// <param name="y">The Vector to process.</param>
/// <returns>Model.</returns>
public override IModel Generate(Matrix x, Vector y)
{
if (this.Descriptor == null)
{
throw new InvalidOperationException("Cannot build naive bayes model without type knowledge!");
}
// create answer probabilities
if (!this.Descriptor.Label.Discrete)
{
throw new InvalidOperationException("Need to use regression for non-discrete labels!");
}
// compute Y probabilities
var statistics = this.GetLabelStats(y);
var root = new Measure { Discrete = true, Label = this.Descriptor.Label.Name, Probabilities = statistics };
// collect feature ranges
var features = this.GetBaseConditionals(x);
// compute conditional counts
for (var i = 0; i < y.Length; i++)
{
var stat = statistics.Where(s => s.X.Min == y[i]).First();
if (stat.Conditionals == null)
{
stat.Conditionals = this.CloneMeasure(features);
}
for (var j = 0; j < x.Cols; j++)
{
var s = stat.Conditionals[j];
s.Increment(x[i, j]);
}
}
// normalize into probabilities
for (var i = 0; i < statistics.Length; i++)
{
var cond = statistics[i];
for (var j = 0; j < cond.Conditionals.Length; j++)
{
cond.Conditionals[j].Normalize();
}
}
return new NaiveBayesModel { Descriptor = this.Descriptor, Root = root };
}
/// <summary>Makes a deep copy of this object.</summary>
/// <returns>A copy of this object.</returns>
public Measure Clone()
{
var m = new Measure { Label = this.Label, Discrete = this.Discrete };
if (this.Probabilities != null && this.Probabilities.Length > 0)
{
m.Probabilities = new Statistic[this.Probabilities.Length];
for (var i = 0; i < m.Probabilities.Length; i++)
{
m.Probabilities[i] = this.Probabilities[i].Clone();
}
}
return m;
}
/// <summary>Makes a deep copy of this object.</summary>
/// <returns>A copy of this object.</returns>
public Measure Clone()
{
var m = new Measure
{
Label = Label,
Discrete = Discrete
};
if (Probabilities != null && Probabilities.Length > 0)
{
m.Probabilities = new Statistic[Probabilities.Length];
for (int i = 0; i < m.Probabilities.Length; i++)
m.Probabilities[i] = Probabilities[i].Clone();
}
return m;
}
/// <summary>Gets base conditionals.</summary>
/// <param name="x">The Matrix to process.</param>
/// <returns>An array of measure.</returns>
private Measure[] GetBaseConditionals(Matrix x)
{
Measure[] features = new Measure[x.Cols];
for (int i = 0; i < features.Length; i++)
{
Property p = Descriptor.At(i);
var f = new Measure
{
Discrete = p.Discrete,
Label = Descriptor.ColumnAt(i),
};
IEnumerable<Statistic> fstats;
if (f.Discrete)
fstats = x[i, VectorType.Col].Distinct().OrderBy(d => d)
.Select(d => Statistic.Make(p.Convert(d).ToString(), d, 1));
else
fstats = x[i, VectorType.Col].Segment(Width)
.Select(d => Statistic.Make(f.Label, d, 1));
f.Probabilities = fstats.ToArray();
features[i] = f;
}
return features;
}
private void LabelIds(Measure m)
{
m.Id = ++_vertexId;
if (m.Probabilities != null)
{
foreach (var s in m.Probabilities)
s.Id = ++_vertexId;
foreach (var s in m.Probabilities)
if (s.Conditionals != null)
foreach (var measure in s.Conditionals)
LabelIds(measure);
}
}