LARA: A Light, Anti‑Overfitting Retraining Method for Unsupervised Time‑Series Anomaly Detection

Introduction

Recently, a paper titled “LARA: A Light and Anti‑overfitting Retraining Approach for Unsupervised Time Series Anomaly Detection”, led by Alibaba Cloud’s big‑data engineering team in collaboration with Zhejiang University, was accepted by WWW2024. The method addresses the problem of cloud services whose normal patterns evolve over time while early‑stage observations are insufficient for model training. LARA can achieve state‑of‑the‑art detection accuracy using only one sample containing 40 time slices for retraining.

Background

Existing solutions such as transfer learning, meta‑learning, and signal‑processing‑based methods either ignore the temporal relationship among historical normal modes or incur excessive computational cost, making them unsuitable for the current problem.

Challenges

Frequent retraining in a changing cloud environment brings two main issues: (1) over‑fitting due to scarce new observations at the beginning of a distribution shift, and (2) huge training overhead.

Breakthrough

To solve these problems, LARA introduces a “rumination” module that uses the old model to recover historical data similar to new observations, jointly estimating a hidden state z for each new point. Linear mapping functions M_z and M_x transform old hidden states and reconstructed data into estimates for the current distribution, minimizing mapping error. A corresponding convex loss design guarantees a unique global optimum, fast convergence, and reduced retraining cost.

Application

LARA has been deployed in Alibaba’s Feitian Big‑Data AI Governance Platform (ABM) anomaly‑detection service, cutting training overhead and enabling timely detection of anomalies.

Paper title: LARA: A Light and Anti‑overfitting Retraining Approach for Unsupervised Time Series Anomaly Detection

Authors: Chen Feiyi, Qin Zhen, Zhou Mengchu, Zhang Yingying, Deng Shuiguang, Fan Lunting, Pang Guansong, Wen Qingsong

PDF: https://arxiv.org/abs/2310.05668