一种竞争自编码器的半监督异常检测方法

Semi-supervised Anomaly Detection Method for Competitive Autoencoders

大多传统的无监督和有监督算法在异常检测上需要预先定义阈值,但很难找到最佳阈值,为解决这一问题,提出一种无阈值的半监督竞争学习模型用于时间序列的异常检测.该模型使用长短期记忆神经网络(LSTM)设计了一种竞争机制的自动编码器,有一个编码器和两个解码器,使用正常数据和未标记数据进行训练,正常数据用来训练正常解码器,对于未标记数据送入两个解码器当中,通过重构误差的大小来选择解码器.该结构具有良好的判别性能,不需要对其设定阈值.实验结果表明,该模型在对异常样本的召回率和F1分数都高于传统的自编码器模型.所以,该半监督竞争学习模型在时间序列的异常检测上是有效的.

Most of the traditional unsupervised and supervised algorithms for anomaly detection need to define the threshold in advance,but it is difficult to find the optimal threshold.To solve this problem,a semi-supervised competitive learning model without threshold is proposed for anomaly detection in time series.The model using LSTM(Long Short Term Memory Network)designed a competition mechanism of the autoencoder,there are a encoder and two decoders,using normal data and unlabelled data for training.The normal data is used train the normal decoder;The unlabelled data is divided into two decoder,through compare reconstruction error to choose the decoder.The structure has good performance,and we needn′t to set a threshold.Experimental results show that the model has higher recall rate and F1-score than the traditional auto-encoder model.Therefore,the semi-supervised competitive learning model is effective in anomaly detection of time series.