[Paper Review] SimAD: A Simple Dissimilarity-based Approach for Time Series Anomaly Detection

Zhijie Zhong, Zhiwen Yu+6 more|arXiv (Cornell University)|May 18, 2024

Anomaly Detection Techniques and Applications4 citations

TL;DR

SimAD proposes a simple yet effective dissimilarity-based approach for time series anomaly detection that leverages an extended temporal receptive field via patch-based feature extraction, an EmbedPatch encoder for robust normal pattern representation, and a ContrastFusion module to amplify distributional differences between normal and abnormal data. It achieves state-of-the-art performance across seven real-world datasets, with relative improvements of up to 19.85% on F1 and 77.79% on NAff-F1, while introducing two new evaluation metrics, UAff and NAff, for fairer performance assessment.

ABSTRACT

Despite the prevalence of reconstruction-based deep learning methods, time series anomaly detection remains a tremendous challenge. Existing approaches often struggle with limited temporal contexts, insufficient representation of normal patterns, and flawed evaluation metrics, all of which hinder their effectiveness in detecting anomalous behavior. To address these issues, we introduce a $ extbf{Sim}$ple dissimilarity-based approach for time series $ extbf{A}$nomaly $ extbf{D}$etection, referred to as $ extbf{SimAD}$. Specifically, SimAD first incorporates a patching-based feature extractor capable of processing extended temporal windows and employs the EmbedPatch encoder to fully integrate normal behavioral patterns. Second, we design an innovative ContrastFusion module in SimAD, which strengthens the robustness of anomaly detection by highlighting the distributional differences between normal and abnormal data. Third, we introduce two robust enhanced evaluation metrics, Unbiased Affiliation (UAff) and Normalized Affiliation (NAff), designed to overcome the limitations of existing metrics by providing better distinctiveness and semantic clarity. The reliability of these two metrics has been demonstrated by both theoretical and experimental analyses. Experiments conducted on seven diverse time series datasets clearly demonstrate SimAD's superior performance compared to state-of-the-art methods, achieving relative improvements of $ extbf{19.85%}$ on F1, $ extbf{4.44%}$ on Aff-F1, $ extbf{77.79%}$ on NAff-F1, and $ extbf{9.69%}$ on AUC on six multivariate datasets. Code and pre-trained models are available at https://github.com/EmorZz1G/SimAD.

Motivation & Objective

• Address the limitations of reconstruction-based deep learning methods in time series anomaly detection, particularly their poor generalization and inadequate modeling of long-term temporal dependencies.
• Overcome the restricted temporal context in existing models by enabling longer effective receptive fields without increasing model complexity.
• Improve representation learning of normal patterns through a novel EmbedPatch encoder that integrates prototype-based attention mechanisms.
• Enhance anomaly discrimination by introducing a ContrastFusion module that amplifies statistical dissimilarities between normal and abnormal distributions.
• Propose and validate two new evaluation metrics, UAff and NAff, to address flaws in existing metrics and ensure fairer benchmarking of anomaly detection models.

Proposed method

• Divide long time series into non-overlapping patches to extend the effective temporal receptive field while maintaining computational efficiency.
• Integrate a prototype-based attention mechanism in each transformer layer to improve feature expressiveness and model capacity without increasing parameter count.
• Employ the EmbedPatch encoder to learn comprehensive, high-level representations of normal behavior by fusing patch embeddings with learned prototypes.
• Design the ContrastFusion module to explicitly enhance the divergence between the distributions of normal and anomalous samples during training.
• Apply asymmetric optimization to focus model learning on abnormal samples, improving sensitivity to rare anomalies.
• Use patch-based cosine similarity as the anomaly score, which better captures long-duration anomalies compared to standard L2 reconstruction loss.

Experimental results

Research questions

• RQ1 Can a dissimilarity-based approach outperform reconstruction-based methods in time series anomaly detection by focusing on distributional differences rather than reconstruction error?
• RQ2 To what extent can patch-based feature extraction and prototype-enhanced attention improve the modeling of long-term temporal dependencies in time series?
• RQ3 How effective is the ContrastFusion module in amplifying the discriminative power between normal and anomalous patterns?
• RQ4 Do the proposed evaluation metrics, UAff and NAff, provide a more reliable and theoretically sound basis for comparing anomaly detection models than existing metrics?
• RQ5 Can the proposed method generalize across diverse multivariate and univariate time series datasets with varying anomaly characteristics?

Key findings

• SimAD achieves a relative improvement of 19.85% on F1 score and 4.44% on Aff-F1 across six multivariate time series datasets compared to state-of-the-art methods.
• On the NAff-F1 metric, SimAD shows a significant 77.79% relative improvement, demonstrating its superior ability to detect anomalies in imbalanced settings.
• The ablation study confirms that removing the EmbedPatch module causes the largest performance drop (on F1), highlighting its critical role in learning meaningful temporal representations.
• The ContrastFusion module contributes substantially to performance, with a 7.88% drop in F1 when removed, indicating its importance in enhancing discrimination between normal and anomalous patterns.
• The data augmentation module improves generalization, reducing UAff-F1 drop by 43.69% when removed, showing its value in robustness to distributional shifts.
• Visualization results on the SWaT dataset confirm that SimAD effectively detects both short- and long-duration anomalies, outperforming AnomTrans and L2-loss-based baselines in segment anomaly detection.