Hybrid approach for Anomaly Detection in Time Series Data
Zeineb Ghrib, Rakia Jaziri, Rim Romdhane
Abstract
Anomaly detection is an active research field which attracts the attention of many business and research actors. It has led to several research projects depending on the nature of the data, the availability of labels on normality, and domains of application that are diverse such as fraud detection, medical domains, cloud monitoring or network intrusions detection, etc. However, dealing with effective anomaly detection for complex and high-dimensional time series data remains a challenging task. In this work, we propose hybrid approach composed of an LSTM Autoencoder trained on normal records to learn efficient normal sequence representations combined with an SVM classifier for anomaly detection. Experimental results show that by encoding time series via a pretrained LSTM encoder allows efficient representation of data so that we can accurately detect abnormal records. In fact, the encoded representation reduces significantly the correlations between normal and abnormal records and allows us to have an efficient latent data representation that separates consistently the two classes. The proposed hybrid approach outperforms state-of-the art approaches [1], [2], [3], [4].