Added Gaussian Multivariate Distribution algorithm

README.txt

@@ -12,6 +12,7 @@ package ml
 	- Logistic Regression
 	- Neural Networks
 	- Collaborative Filtering
+	- Gaussian Multivariate Distribution for anomaly detection systems
 
     Is implemented too the fmincg function in order to calculate the optimal
     theta configuration to reduce the cost value for all the implemented
@@ -152,6 +153,32 @@ type DataSet interface {
 
 
 
+type MultGaussianDist struct {
+    X      [][]float64 // Data to train the gussian distribution
+    Sigma2 []float64
+    Mu     []float64 // Medians for each feature
+}
+    Anomaly detection implementation using Multivariate Gaussian
+    Distribution:
+    http://en.wikipedia.org/wiki/Multivariate_normal_distribution#Density_function
+
+
+func MultVarGaussianDistLoadFromFile(filePath string) (gd *MultGaussianDist)
+    Creates a MultGaussianDist object from the content of a CSV file space
+    sepparate where each line is a sample and each column a feature
+
+
+func (gd *MultGaussianDist) CalculateMuSigma()
+    To be called before GetProbability method in order to calculate the
+    medians and sigmas params for all the training set
+
+func (gd *MultGaussianDist) GetProbability(data [][]float64) (p []float64)
+    Returns the probability of anomaly for each data, each row of data is a
+    sample to study and each colum a featurea, determinate an epsilon and
+    when p(x) < epsilon, you may have an anomaly, you can use
+    SelectThreshold in in order to calculate the best epsilon
+
+
 type NeuralNet struct {
     // Training set of values for each feature, the first dimension are the test cases
     X [][]float64

ml.go

@@ -6,6 +6,7 @@
 //    - Logistic Regression
 //    - Neural Networks
 //    - Collaborative Filtering
+//    - Gaussian Multivariate Distribution for anomaly detection systems
 //
 // Is implemented too the fmincg function in order to calculate the optimal
 // theta configuration to reduce the cost value for all the implemented solutions.

multivariate_gaussian_normal_distribution.go

@@ -0,0 +1,99 @@
+package ml
+
+import (
+	"io/ioutil"
+	"math"
+	"strconv"
+	"strings"
+)
+
+// Anomaly detection implementation using Multivariate Gaussian Distribution:
+// http://en.wikipedia.org/wiki/Multivariate_normal_distribution#Density_function
+type MultGaussianDist struct {
+	X      [][]float64 // Data to train the gussian distribution
+	Sigma2 []float64
+	Mu     []float64 // Medians for each feature
+}
+
+// To be called before GetProbability method in order to calculate the medians
+// and sigmas params for all the training set
+func (gd *MultGaussianDist) CalculateMuSigma() {
+	gd.Mu = make([]float64, len(gd.X[0]))
+
+	for _, x := range gd.X {
+		for i, val := range x {
+			gd.Mu[i] += val
+		}
+	}
+	for i := 0; i < len(gd.X[0]); i++ {
+		gd.Mu[i] /= float64(len(gd.X))
+	}
+
+	gd.Sigma2 = make([]float64, len(gd.X[0]))
+	for _, x := range gd.X {
+		for i, val := range x {
+			gd.Sigma2[i] += math.Pow(val-gd.Mu[i], 2)
+		}
+	}
+	for i := 0; i < len(gd.X[0]); i++ {
+		gd.Sigma2[i] /= float64(len(gd.X))
+	}
+}
+
+// Returns the probability of anomaly for each data, each row of data is a
+// sample to study and each colum a featurea, determinate an epsilon and when
+// p(x) < epsilon, you may have an anomaly, you can use SelectThreshold in
+// in order to calculate the best epsilon
+func (gd *MultGaussianDist) GetProbability(data [][]float64) (p []float64) {
+	for i := 0; i < len(data); i++ {
+		for c := 0; c < len(data[0]); c++ {
+			data[i][c] -= gd.Mu[c]
+		}
+	}
+
+	detSig := 1.0
+	for i := 0; i < len(gd.Sigma2); i++ {
+		detSig *= gd.Sigma2[i]
+	}
+
+	base := math.Pow(2*math.Pi, -(float64(len(data[0])))/2) * math.Pow(detSig, -0.5)
+	p = make([]float64, len(data))
+	for i := 0; i < len(data); i++ {
+		for c := 0; c < len(data[0]); c++ {
+			p[i] += -0.5 * data[i][c] * data[i][c] * (1 / gd.Sigma2[c])
+		}
+		p[i] = base * math.Pow(math.E, p[i])
+	}
+	return
+}
+
+// Creates a MultGaussianDist object from the content of a CSV file space
+// sepparate where each line is a sample and each column a feature
+func MultVarGaussianDistLoadFromFile(filePath string) (gd *MultGaussianDist) {
+	strInfo, err := ioutil.ReadFile(filePath)
+	if err != nil {
+		panic(err)
+	}
+
+	trainingData := strings.Split(string(strInfo), "\n")
+	gd = &MultGaussianDist{
+		X: make([][]float64, len(trainingData)-1),
+	}
+	for i, line := range trainingData {
+		if line == "" {
+			break
+		}
+
+		parts := strings.Split(line, " ")
+		gd.X[i] = make([]float64, len(parts))
+		for c, value := range strings.Split(line, " ") {
+			floatVal, err := strconv.ParseFloat(value, 64)
+			if err != nil {
+				panic(err)
+			}
+			gd.X[i][c] = floatVal
+		}
+	}
+
+	return
+}

multivariate_gaussian_normal_distribution_test.go

@@ -0,0 +1,32 @@
+package ml
+
+import (
+	"fmt"
+	"testing"
+)
+
+func TestMultivariateGaussianDistreibution(t *testing.T) {
+	fmt.Println("Testing Multivariate Gaussian Distribution implementation...")
+	gd := MultVarGaussianDistLoadFromFile(
+		"test_data/multivarite_gaussian_normal_distribution.dat")
+	gd.CalculateMuSigma()
+	probs := gd.GetProbability([][]float64{
+		[]float64{13.0468, 14.7412},
+		[]float64{13.4085, 13.7633},
+		[]float64{14.1959, 15.8532},
+		[]float64{14.9147, 16.1743},
+	})
+
+	expectedRes := []float64{
+		0.06470823722117165,
+		0.050304834321416796,
+		0.07244977886910747,
+		0.0503144083695188,
+	}
+
+	for i, res := range expectedRes {
+		if res != probs[i] {
+			t.Error("The expected value is:", res, "but the obtained was:", probs[i])
+		}
+	}
+}