def __init__(self,
anomalies,
unlabel,
classifer,
return_proba=True,
n_clusters='auto',
cluster_algo='kmeans',
contamination=0.02,
theta=0.85,
alpha='auto',
beta='auto',
random_state=2018):
scaler = StandardScaler()
self.anomalies = scaler.fit_transform(anomalies) # 对应于已知的正样本集(P集)
self.unlabel = scaler.fit_transform(unlabel) # 对应于无标签样本集(U集)
self.n_clusters = n_clusters # 聚类簇数可以预先指定，也可以由get_cluster_centers自动确定最佳聚类簇数
self.classifer = classifer # 选取的分类器
self.return_proba = return_proba # 布尔型参数，是否返回样本取正的后验概率
self.cluster_algo = cluster_algo # 聚类算法，可以选取'spectral'、'birch'、'dbscan'、'kmeans'，默认取'kmeans'
self.contamination = contamination # contamination为预估的U中异常样本(即正样本)比例
self.theta = theta # isolation_score、similarity_score的加权系数分别为theta、1-theta
self.alpha = alpha # 论文默认取已知异常样本(P集)的total_score均值，作为无标签样本是否为potential anomalies的阈值
self.beta = beta # 判定无标签样本是否为reliable normal的阈值
self.random_state = random_state
self.centers = get_cluster_centers(self.anomalies, self.n_clusters,
self.cluster_algo) # 返回聚类的簇中心
def __init__(self, anomalies, unlabel, classifer, cluster_algo='kmeans', n_clusters='auto',
contamination=0.01, theta=0.85, alpha='auto', beta='auto', return_proba=False,
random_state=2018):
'''
:param anomalies: Observed anomaly data sets
:param unlabel: Unlabeled data sets.
:param classifer: A Classifer fitting weighted samples and labels to predict unlabel samples.
:param cluster_algo: str, {'kmeans'、'spectral'、'birch'、'dbscan'}, default = 'kmeans'
Clustering algorithm for clustering anomaly samples.
:param n_clusters: int, default=5
The number of clusters to form as well as the number of centroids to generate.
:param contamination : float, range (0, 0.5).
The proportion of outliers in the data set.
:param theta : float, range [0, 1].
The weights of isolation_score and similarity_score are theta and 1-theta respectively.
:param alpha : float, should be positive number, default = mean value of anomalies's score
Threshold value for determining unlabel sample as potential anomaly
:param beta : float, should be positive number
Threshold value for determining unlabel sample as reliable normal sample
:param return_proba : bool, default=False
Whether return the predicted probability for positive(anomaly) class for each sample.
Need classifer to provide predict_proba method.
'''
dataset_scaled = StandardScaler().fit_transform(np.r_[anomalies, unlabel])
self.anomalies = dataset_scaled[:len(anomalies), :]
self.unlabel = dataset_scaled[len(anomalies):, :]
self.contamination = contamination
self.classifer = classifer
self.n_clusters = n_clusters
self.cluster_algo = cluster_algo
self.theta = theta
self.alpha = alpha
self.beta = beta
self.return_proba = return_proba
self.random_state = random_state
self.centers, self.cluster_score = get_cluster_centers(self.anomalies, self.n_clusters, self.cluster_algo)
def __init__(self,
anomalies,
unlabel,
classifer,
cluster_algo='kmeans',
n_clusters='auto',
kernel='rbf',
verbose=3,
contamination=0.01,
theta=0.85,
alpha='auto',
beta='auto',
return_proba=False,
random_state=2018):
'''
:param anomalies: Observed anomaly datasets.
:param unlabel: Unlabeled datasets.
:param classifer: A Classifer fitting weighted samples and labels to predict unlabel samples.
:param cluster_algo: str, {'kmeans'、'spectral'、'birch'、'dbscan'}, default = 'kmeans'
Clustering algorithm for clustering anomaly samples.
:param n_clusters: int, default=5
The number of clusters to form as well as the number of centroids to generate.
:param kernel: str, default='rbf'.
'linear' | 'poly' | 'rbf' | 'sigmoid' | 'cosine' | 'precomputed' kernel.
:param verbose: int, default=3, Verbosity mode. the higher, the less messages.
KernelPCA is time-consuming, and the verbose parameter helps to check the progress
of the reconstruction. If verbose = m, information is printed every m rounds.
:param contamination : float, range (0, 0.5). The proportion of outliers in the data set.
:param theta : float, range [0, 1].
The weights of isolation_score and similarity_score are theta and 1-theta respectively.
:param alpha : float, should be positive number, default = mean value of anomalies's score
Threshold value for determining unlabel sample as potential anomaly
:param beta : float, should be positive number
Threshold value for determining unlabel sample as reliable normal sample
:param return_proba : bool, default=False
Whether return the predicted probability for positive(anomaly) class for each sample.
Need classifer to provide predict_proba method.
'''
self.dataset = StandardScaler().fit_transform(np.r_[anomalies,
unlabel])
self.anomalies = self.dataset[:len(anomalies), :]
self.unlabel = self.dataset[len(anomalies):, :]
self.contamination = contamination
self.verbose = verbose
self.kernel = kernel
self.classifer = classifer
self.n_clusters = n_clusters
self.cluster_algo = cluster_algo
self.theta = theta
self.alpha = alpha
self.beta = beta
self.return_proba = return_proba
self.random_state = random_state
self.centers, self.cluster_score = get_cluster_centers(
self.anomalies, self.n_clusters, self.cluster_algo)