标签：
K近邻应用-异常检测应用
原理：
根据数据样本进行KMeans机器学习模型的建立，获取簇心点，以簇为单位，离簇心最远的第五个点的距离为阈值，大于这个值的为异常点，即获得数据异常。
如图：
数据样本：
1,2.43,2.3899999 2,2.38,2.12 3,2.8,2.51 4,2.01,2.69 5,2.71,2.45 6,2.55,2.34
7,2.46,2.31 8,2.27,2.38 9,2.87,2.55 10,2.75,2.07 11,2.3899999,2.6100001 12,2.67,2.31
13,2.68,2.75 14,2.47,2.05 15,2.96,2.66 16,2.08,2.92 17,2.58,2.12 18,2.69,2.72
19,2.29,2.81 20,2.2,2.2 21,2.46,2.87 22,2.66,2.92 23,2.71,2.63 24,2.09,2.99
25,2.33,2.84 26,2.4,2.63 27,2.05,2.27 28,2.59,2.81 29,2.68,2.72 30,2.5,2.29
31,2.63,2.8899999 32,2.35,2.8600001 33,2.74,2.06 34,2.83,2.56 35,2.3600001,2.87 36,2.25,2.32
37,2.99,2.85 38,2.19,2.62 39,2.37,2.19 40,2.37,2.08 41,2.62,2.25 42,2.16,2.56
43,2.08,2.37 44,2.77,2.55 45,2.96,2.85 46,2.52,2.24 47,2.6,2.55 48,2.78,2.14
49,2.76,2.42 50,2.05,2.67 51,2.94,2.82 52,2.52,2.59 53,2.04,2.08 54,2.65,2.03
55,2.32,2.88 56,2.96,2.2 57,2.97,2.28 58,2.01,2.6399999 59,2.58,2.52 60,2.55,2.7
61,2.75,2.19 62,2.28,2.48 63,2.6399999,2.54 64,2.34,2.27 65,2.72,2.23 66,2.5,2.35
67,2.25,2.2 68,2.27,2.91 69,2.8899999,2.88 70,2.76,2.48 71,2.63,2.22 72,2.69,2.33
73,2.9,2.02 74,2.23,2.26 75,2.82,2.87 76,2.57,2.83 77,2.97,2.47 78,2.69,2.54
79,2.6,2.84 80,2.98,2.99 81,2.21,2.3899999 82,2.11,2.46 83,2.54,2.77 84,2.57,2.19
85,2.66,2.77 86,2.4,2.88 87,2.43,2.75 88,2.35,2.05 89,2.68,2.25 90,2.43,2.87
91,2.06,2.05 92,2.8600001,2.6100001 93,2.58,2.75 94,2.91,2.8 95,2.38,2.95 96,2.63,2.58
97,2.82,2.93 98,2.72,2.97 99,2.16,2.55 100,5.46,5.1 101,5.9,5.39 102,5.81,5.91
103,5.92,5.65 104,5.91,5.94 105,5.9,5.91 106,5.7799997,5.66 107,5.76,5.32 108,5.11,5.77
109,5.38,5.46 110,5.63,5.76 111,5.1,5.7200003 112,5.66,5.31 113,5.86,5.6 114,5.46,5.74
115,5.76,5.17 116,5.39,5.24 117,5.33,5.49 118,5.05,5.28 119,5.8,5.63 120,5.0,5.18
121,5.35,5.71 122,5.5299997,5.45 123,5.95,5.04 124,5.17,5.32 125,5.83,5.56 126,5.67,5.55
127,5.63,5.25 128,5.42,5.27 129,5.38,5.57 130,5.39,5.6 131,5.88,5.41 132,5.84,5.38
133,5.95,5.36 134,5.65,5.43 135,5.76,5.05 136,5.65,5.5 137,5.13,5.07 138,5.79,5.87
139,5.87,5.38 140,5.9,5.96 141,5.28,5.05 142,5.8,5.61 143,5.24,5.24 144,5.08,5.35
145,5.38,5.5299997 146,5.4,5.62 147,5.73,5.0 148,5.3,5.1 149,5.34,5.39 150,5.63,5.34
151,5.4,5.29 152,5.23,5.26 153,5.04,5.25 154,5.49,5.83 155,5.89,5.18 156,5.18,5.85
157,5.41,5.67 158,5.81,5.7200003 159,5.62,5.41 160,5.79,5.5 161,5.35,5.94 162,5.31,5.68
163,5.14,5.74 164,5.37,5.59 165,5.19,5.91 166,5.62,5.64 167,5.26,5.38 168,5.74,5.91
169,5.17,5.8 170,5.68,5.13 171,5.67,5.21 172,5.2,5.49 173,5.89,5.87 174,5.8,5.22
175,5.01,5.31 176,5.0,5.28 177,5.95,5.56 178,5.27,5.23 179,5.9,5.74 180,5.21,5.75
181,5.13,5.3 182,5.36,5.0 183,5.21,5.86 184,5.21,5.56 185,5.7799997,5.15 186,5.04,5.4
187,5.52,5.61 188,5.02,5.99 189,5.32,5.04 190,5.81,5.51 191,5.76,5.29 192,5.03,5.62
193,5.08,5.26 194,5.42,5.4 195,5.28,5.04 196,5.2,5.49 197,5.7799997,5.33 198,5.38,5.71
199,5.9700003,5.96 200,8.51,8.93 201,8.43,8.58 202,8.62,8.31 203,8.08,8.52 204,8.31,8.49
205,8.4,8.97 206,8.6,8.74 207,8.96,8.76 208,8.0,8.79 209,8.04,8.0 210,8.71,8.23
211,8.78,8.4 212,8.85,8.34 213,8.04,8.74 214,8.92,8.55 215,8.0,8.9 216,8.24,8.45
217,8.33,8.35 218,8.83,8.94 219,8.23,8.06 220,8.46,8.85 221,8.39,8.59 222,8.7,8.85
223,8.45,8.68 224,8.86,8.74 225,8.11,8.18 226,8.11,8.27 227,8.15,8.35 228,8.99,8.27
229,8.67,8.12 230,8.18,8.92 231,8.58,8.58 232,8.05,8.67 233,8.97,8.11 234,8.76,8.49
235,8.18,8.54 236,8.82,8.64 237,8.74,8.89 238,8.82,8.77 239,8.02,8.33 240,8.77,8.54
241,8.22,8.13 242,8.92,8.35 243,8.71,8.55 244,8.12,8.74 245,8.07,8.96 246,8.71,8.17
247,8.12,8.4 248,8.03,8.92 249,8.99,8.55 250,8.63,8.19 251,8.95,8.82 252,8.25,8.32
253,8.08,8.21 254,8.31,8.94 255,8.87,8.3 256,8.72,8.23 257,8.98,8.88 258,8.48,8.64
259,8.81,8.3 260,8.15,8.07 261,8.36,8.02 262,8.16,8.22 263,8.77,8.44 264,8.51,8.17
265,8.28,8.31 266,8.57,8.47 267,8.95,8.1 268,8.91,8.72 269,8.34,8.64 270,8.07,8.99
271,8.3,8.75 272,8.35,8.75 273,8.9,8.22 274,8.99,8.94 275,8.67,8.37 276,8.27,8.0
277,8.68,8.93 278,8.18,8.45 279,8.25,8.82 280,8.99,8.17 281,8.36,8.17 282,8.64,8.38
283,8.94,8.77 284,8.33,8.71 285,8.23,8.81 286,8.56,8.79 287,8.71,8.89 288,8.09,8.27
289,8.93,8.0 290,8.66,8.23 291,8.35,8.1 292,8.15,8.54 293,8.72,8.03 294,8.64,8.76
295,8.94,8.28 296,8.39,8.87 297,8.01,8.4 298,8.07,8.28 299,8.12,8.65 300,8.65,8.16
数据样本的数据格式为：标号，特征值1，特征值2（没有具体含义，自动生成的数据只为能够简单的说明异常检测是怎么一回事，以及机器学习到底是如何应用在实际生产环境中）
可视化展示：
我们将数据样本投射到可视化环境中的可以看到数据呈现以下图形：
数据被分为3簇，在我们训练模型是K值为3簇。由于数据非常集中，数据量也非常少，同时特征向量为二维特征向量，故投影成平面图形我们一眼可以看出数据分为几簇，当样本数据的特征值很多时，就得靠计算得出K值（这里先不提）
应用代码实践：
//获取样本数据
val rawData = sc.textFile("D:/logdata/kmeans.txt")
//将样本数据转化为模型可操作的向量集
val labelAndData = rawData.map { line =>
val buffer = line.split(‘,‘).toBuffer
val label = buffer.remove(0)
val vector = Vectors.dense(buffer.map(_.toDouble).toArray)
(label, vector)
}
//将样本数据向量集缓存
val data = labelAndData.values.cache()
//建立Kmeans学习模型
val kmeans = new KMeans()
kmeans.setK(3)
//训练数据
val model = kmeans.run(data)
//打印簇心点
model.clusterCenters.foreach(println)
//欧氏距离的计算函数
def distance(a: Vector, b: Vector): Double = {
math.sqrt(a.toArray.zip(b.toArray).map(p => p._1 - p._2).map(d => d * d).sum)
}
//计算向量到模型簇心点的距离
def distToCentroid(datum: Vector, model: KMeansModel) = {
val cluster = model.predict(datum)
val centroid = model.clusterCenters(cluster)
distance(centroid, datum)
}
//计算所有点到簇心点的距离集合
val distances = data.map(datum =>
distToCentroid(datum, model)
)
//获取最大的第五个值为阈值
val threshold = distances.top(5).last
//测试数据获取
val testRawData = sc.textFile("D:/logdata/kmeans")
val testLabelAndData = testRawData.map { line =>
val buffer = line.split(‘,‘).toBuffer
val label = buffer.remove(0)
val vector = Vectors.dense(buffer.map(_.toDouble).toArray)
(label, vector)
}
//将测试数据集缓存
val testData = testLabelAndData.values.cache()
//异常数据集过滤并打印结果
val anomalies=testData.filter { x =>
distToCentroid(x, model) > threshold
}.collect().foreach(println)
计算结果：
[5.525200003000001,5.494100009000001]
[2.522222221212122,2.512020205050505]
[8.483267326732673,8.49178217821782]
异常值：
[6.73,6.58]
[6.62,6.04]
[6.99,6.66]
[6.59,6.38]
[6.42,6.74]
[6.37,6.59]
[6.84,6.03]
[6.84,6.03]
[6.9700003,6.5299997]
[6.03,6.31]
[6.18,6.27]
[6.84,6.81]
[6.3,6.93]
[6.49,6.23]
[6.16,6.67]
[6.56,6.77]
[6.57,6.32]
[6.37,6.55]
[6.68,6.07]
[6.8,6.4]
[6.91,6.44]
标签：
原文地址：http://www.cnblogs.com/gnool/p/5883209.html
