异常检测LOF算法(手码和调库-python)

全称又叫局部异常因子，是一种基于距离的异常点检测算法。
理论链接: LOF算法理论.

• 手码

由于需要搜寻最近的K个点，采用KDTree来加快效率

from scipy import spatial

# items in data_list should be unique 
# K is the number of neighbors
# return LOF score
def getLOF(data_list,K,test_data):
    #if data_list contains test_data ,ignore it 
    K+=1
    
    #establish KDTree
    tree =spatial.KDTree(data_list)
    
    #get KNodes of test_data
    KNodes,distanceList=getKNodesAndDistance(tree,K,test_data)
    
    #get Reach Distance and Density of test
    lrdistance_testData=getReachDistance(KNodes,tree,K,test_data)
    lrdensity_testData=1/(lrdistance_testData/len(KNodes))
    
    #get Density of TestNode's KNodesList 
    totalReachDensity=getReachDensity(KNodes,tree,K)
    
    return (totalReachDensity/lrdensity_testData)/len(KNodes)

# Euclid distance
def getDistance(a,b):
    return math.sqrt(math.pow(a[0]-b[0],2)+math.pow(a[1]-b[1],2))

def getKNodesAndDistance(kdtree,K,test_data):
    #get KNodes of test_data
    list_distance,list_index=kdtree.query(test_data,k=K,p=2)
    KNodes = []
    for temp in list_index:
        KNodes.append(kdtree.data[temp])
    return KNodes,list_distance

def getReachDistance(KNodes,tree,K,test_data):
    totalReachDistance=0
    for temp_Nodes in KNodes:
        KNodes_temp_Nodes,distanceList_temp_Nodes = getKNodesAndDistance(tree,K,temp_Nodes)
        reachDistance_everyKNode=getDistance(temp_Nodes,test_data)
#         可达距离指周围K领域里的点到test点的K距离和直线距离的最大值
#          K距离指距离周围的最近K个点的最大距离
        if(distanceList_temp_Nodes[len(distanceList_temp_Nodes)-1]>reachDistance_everyKNode):
             reachDistance_everyKNode=distanceList_temp_Nodes[len(distanceList_temp_Nodes)-1]
        totalReachDistance+=reachDistance_everyKNode
    return totalReachDistance

def getReachDensity(KNodes,tree,K):
    totalReachDensity = 0
    for temp_Node in KNodes:
        KNodes_temp_Node,distance = getKNodesAndDistance(tree,K,temp_Node)
        
        reachDistacne_temp_Node=getReachDistance(KNodes_temp_Node,tree,K,temp_Node)
        
        reachDensity_temp_Node=1/(reachDistacne_temp_Node/len(KNodes_temp_Node))
        
        totalReachDensity+=reachDensity_temp_Node
    
    return totalReachDensity

• 调库

from sklearn.neighbors import LocalOutlierFactor as LOF

# y得到的是-1、1列表，-1表示为异常值
model = LOF(n_neighbors=5, contamination=0.1,novelty=True,n_jobs=-1)    #定义一个LOF模型，异常比例是10%
model.fit(list(set(list_fortest)))
y = model._predict(list(set(list_fortest)))