Taylorboost: reinterpreting taylor expansion while boosting anomaly detection

Abstract:
Anomaly detection and fault localization are key functions in telecom network management systems. Network devices (i.e. entities) such as routers, switches, transmitters, and so on are typically monitored with multivariate time series, the detection of anomalies being critical for an entity's service quality management. Nevertheless, given the complexity of multivariate time series, detecting anomalies is still challenging. These two functions can be attributed to the same problem, the problem of anomaly detection for multivariate time series. We propose Taylor features, which are filtered using the Boosting algorithm and then transformed into a hierarchical equivalent representation. Further, we use stochastic RNN to capture temporal dependencies of sequences and FCM to model the relationship among variables. Finally, it comes to the TaylorBoost model. The experiments are carried out on a new server machine dataset from an Internet company. TaylorBoost outperforms other baseline methods with an overall precision of 0.90.