SimAD: A Simple Dissimilarity-Based Approach for Time-Series Anomaly Detection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time-series anomaly detection (TSAD) faces challenges with limited temporal context and inadequate normal pattern representation.
• Existing deep learning methods often struggle with detecting anomalous behavior effectively.
• Current evaluation metrics for TSAD may lack distinctiveness and semantic clarity.

Purpose of the Study:

• To introduce SimAD, a simple dissimilarity-based approach for robust time-series anomaly detection.
• To address limitations in temporal context, normal pattern representation, and evaluation metrics in TSAD.
• To improve the accuracy and reliability of anomaly detection in time-series data.

Main Methods:

• SimAD utilizes a patching-based feature extractor for extended temporal windows and the EmbedPatch encoder for integrating normal patterns.
• A ContrastFusion module enhances robustness by highlighting distributional differences between normal and abnormal data.
• Introduced two novel evaluation metrics: unbiased affiliation (UAff) and normalized affiliation (NAff).

Main Results:

• SimAD demonstrated superior performance over state-of-the-art methods on seven diverse time-series datasets.
• Achieved significant relative improvements: 19.85% on F1, 4.44% on Aff-F1, 77.79% on NAff-F1, and 9.69% on AUC on multivariate datasets.
• The proposed UAff and NAff metrics proved reliable and effective in evaluating TSAD performance.

Conclusions:

• SimAD offers a simple yet effective solution for time-series anomaly detection.
• The method successfully addresses key challenges in existing TSAD approaches.
• The developed evaluation metrics provide clearer and more robust assessments of anomaly detection systems.