UNIANO: robust and efficient anomaly consensus in time series sensitive to cross-correlated anomaly profiles
Leonor Silva, Héléna Galhardas, Vasco M. Manquinho, Rui Henriques · Society for Industrial and Applied Mathematics eBooks · 2021
Time series anomaly detection is an active research area, combining dozens of state-of-the-art methods that place heterogeneous views on what is an anomaly.This diversity of views -local and global, point and segment, univariate and multivariate, context-free and context-aware anomalies -is associated with moderate-to-high output divergences between methods.As a result, the user is faced with the difficult and laborious task of selecting the most appropriate methods and identifying cross-method consensus in an attempt to optimize recall and precision.Despite the relevance of establishing agreement criteria, existing principles are scarce and suffer from major problems: 1) show biases towards methods with correlated/redundant anomaly profiles; 2) depend on anomaly score thresholding; 3) prevent online detection; and 4) offer consensus not subjected to sound statistical testing.This work proposes UNIANO (UNIfied ANOmaly), an approach that combines simple yet effective empirical multivariate distribution statistics to address these drawbacks, guaranteeing a parameter-free and statistically robust integration of heterogeneous anomaly views.In this context, anomalies detected by less prevalent and concordant anomaly profiles, such as context-aware profiles in the presence of complementary variables, are not undervalued.Given a n-length time series and m views, UNIANO is aided by adequate data structures to achieve O(n log m 2 n) training time and linear O(m) testing-and-updating time.The gathered results confirm the relevance of the proposed approach.