在 Accord.Net 中如何使用 One-Class SVM 进行异常检测？

Question

我正在尝试通过在 Accord.Net 中使用 OneclassSupportVectorLearning 来实现异常检测。我在训练过程中遇到了 NullReference 错误。下面是我在测试中的示例代码。 感谢是否有人可以帮助我解决这个问题。

 double[][] inputs =
 {
    new double[] { 0, 1, 1, 0 }, //  0 
    new double[] { 0, 1, 0, 0 }, //  0
    new double[] { 0, 0, 1, 0 }, //  0
    new double[] { 0, 1, 1, 0 }, //  0
    new double[] { 0, 1, 0, 0 }, //  0
 };
 var oteacher = new OneclassSupportVectorLearning<ChiSquare,double[]>();
 var k = oteacher.Learn(inputs); //NullReference error occur here. 

编辑------------------------------ ----------------------

根据 Jstreet 的评论，尝试下面的代码，但它适用于 2-dim，但在更高维度上会失败。

static void Main(string[] args)
{
Random r = new Random(DateTime.Now.Millisecond);

int size = 1000;
int min = 45;
int max = 55;

double[][] inputs = new double[size][];

for (int i = 0; i < size; i++)
{
    double[] d = new double[] { r.Next(min,max), r.Next(min,max), r.Next(min,max), r.Next(min,max) };
    inputs[i] = d;
}

var oteacher = new OneclassSupportVectorLearning<ChiSquare>();
var k = oteacher.Learn(inputs);

double[][] test =
 {
    // normal
    new double[] { 50, 53 , 50, 50}, 
    new double[] { 49, 52 , 50, 50},
    new double[] { 48, 51 , 50, 50},
    new double[] { 47, 52 , 50, 50},
    new double[] { 46, 53 , 50, 50},
    // anomalies
    new double[] { 50, 70, 70, 70 }, 
    new double[] { 51, 69, 70, 70 },
    new double[] { 52, 68, 70, 70 },
    new double[] { 53, 67, 70, 70 },
    new double[] { 54, 66, 70, 70 },
 };

foreach (double[] d in test)
{
    if (k.Decide(d) == true)
        Console.WriteLine(" OK = {0}, {1}, {2}, {3}", d[0], d[1], d[2], d[3]);
    else Console.WriteLine(" Anomaly = {0}, {1}, {2}, {3}", d[0], d[1], d[2], d[3]);
}

Console.ReadLine();

}

Answer 1

我建议您尝试使用二维数据集，以便将结果可视化并感受一下：

    static void Main(string[] args)
    {
        Random r = new Random(DateTime.Now.Millisecond);

        int size = 100;
        int min = 45;
        int max = 55;

        double[][] inputs = new double[size][];

        for (int i = 0; i < size; i++)
        {
            double[] d = new double[] { r.Next(min,max), r.Next(min,max) };
            inputs[i] = d;
        }

        var oteacher = new OneclassSupportVectorLearning<ChiSquare>();
        var k = oteacher.Learn(inputs);

        double[][] test =
         {
            // normal
            new double[] { 50, 53 }, 
            new double[] { 49, 52 },
            new double[] { 48, 51 },
            new double[] { 47, 52 },
            new double[] { 46, 53 },
            // anomalies
            new double[] { 50, 70 }, 
            new double[] { 51, 69 },
            new double[] { 52, 68 },
            new double[] { 53, 67 },
            new double[] { 54, 66 },
         };

        foreach (double[] d in test)
        {
            if (k.Decide(d) == true)
                Console.WriteLine(" OK = {0}, {1}", d[0], d[1]);
            else Console.WriteLine(" Anomaly = {0}, {1}", d[0], d[1]);
        }

        Console.ReadLine();
    }

此示例代码生成以下输出：

 OK = 50, 53 
 OK = 49, 52 
 OK = 48, 51 
 OK = 47, 52 
 OK = 46, 53 
 Anomaly = 50, 70 
 Anomaly = 51, 69 
 Anomaly = 52, 68 
 Anomaly = 53, 67 
 Anomaly = 54, 66

这是相同结果的图形视图：

---

编辑：就像我说的，这需要一些实验。这是我对 4 维输入数据集的结果。请注意，我减少了每个维度的可变性并保持相同的输入大小，100。

    static void Main(string[] args)
    {
        Random r = new Random(DateTime.Now.Millisecond);

        int size = 100;
        int min = 45;
        int max = 50;
        int min2 = 60;
        int max2 = 65;

        double[][] inputs = new double[size][];

        for (int i = 0; i < size; i++)
        {
            double[] d = new double[] { r.Next(min, max), r.Next(min, max), r.Next(min, max), r.Next(min, max) };
            inputs[i] = d;
        }

        var oteacher = new OneclassSupportVectorLearning<ChiSquare>();
        var k = oteacher.Learn(inputs);

        double[][] test =
         {
            // normal
            new double[] {  r.Next(min, max),  r.Next(min, max), r.Next(min, max),  r.Next(min, max) },
            new double[] {  r.Next(min, max),  r.Next(min, max), r.Next(min, max),  r.Next(min, max) },
            new double[] {  r.Next(min, max),  r.Next(min, max), r.Next(min, max),  r.Next(min, max) },
            new double[] {  r.Next(min, max),  r.Next(min, max), r.Next(min, max),  r.Next(min, max) },
            new double[] {  r.Next(min, max),  r.Next(min, max), r.Next(min, max),  r.Next(min, max) },
            // anomalies
            new double[] {  r.Next(min2, max2),  r.Next(min2, max2), r.Next(min2, max2),  r.Next(min2, max2) },
            new double[] {  r.Next(min2, max2),  r.Next(min2, max2), r.Next(min2, max2),  r.Next(min2, max2) },
            new double[] {  r.Next(min2, max2),  r.Next(min2, max2), r.Next(min2, max2),  r.Next(min2, max2) },
            new double[] {  r.Next(min2, max2),  r.Next(min2, max2), r.Next(min2, max2),  r.Next(min2, max2) },
            new double[] {  r.Next(min2, max2),  r.Next(min2, max2), r.Next(min2, max2),  r.Next(min2, max2) },
         };

        foreach (double[] d in test)
        {
            if (k.Decide(d) == true)
                Console.WriteLine("OK = {0}, {1}, {2}, {3}", d[0], d[1], d[2], d[3]);
            else Console.WriteLine("Anomaly = {0}, {1}, {2}, {3}", d[0], d[1], d[2], d[3]);
        }

        Console.ReadLine();
    }

结果：

OK = 49, 46, 47, 49
OK = 49, 45, 45, 47
OK = 45, 45, 46, 47
OK = 47, 49, 47, 48
OK = 45, 45, 47, 48
Anomaly = 62, 60, 61, 63
Anomaly = 61, 63, 63, 64
Anomaly = 64, 60, 60, 64
Anomaly = 61, 64, 63, 63
Anomaly = 62, 60, 62, 62