Multivariate Time Series Anomaly Detection by Capturing Coarse-Grained Intra- and Inter-Variate Dependencies

Yongzheng Xie, Hongyu Zhang, Muhammad Ali Babar · 2025

Read Multivariate Time Series Anomaly Detection by Capturing Coarse-Grained Intra- and Inter-Variate Dependencies by Yongzheng Xie, Hongyu Zhang, Muhammad Ali Babar (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-16364.txt.

Introduction

Yongzheng Xie, Hongyu Zhang, and Muhammad Ali Babar build a detector for faults in systems watched by many sensors at once. Only normal data can be used for training, since failures are rare and labeling is costly. Most detectors learn each sensor's rhythm and the links between sensors at a very fine grain, which the authors argue misses the real patterns.

Their method, MtsCID, zooms out: one network reads coarse patches of each sensor's history, another studies coarse links between sensors against wavy reference patterns. Across seven public datasets, the coarser view matched or beat the best existing detectors. The team frames grain size as the overlooked dial in anomaly detection.

Catalog record

CREATORYongzheng Xie, Hongyu Zhang, Muhammad Ali Babar
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-16364

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Multivariate Time Series Anomaly Detection by Capturing Coarse-Grained Intra- and Inter-Variate Dependencies Yongzheng Xie, Hongyu Zhang, Muhammad Ali Babar Abstract . Multivariate time series anomaly detection is essential for failure management in web application operations, as it directly influences the effectiveness and timeliness of implementing remedial or preventive measures. This task is often framed as a semi-supervised learning problem, where only normal data are available for model training, primarily due to the labor-intensive nature of data labeling and the scarcity of anomalous data. Existing semi-supervised methods often detect anomalies by capturing intra-variate temporal dependencies and/or inter-variate relationships to learn normal patterns, flagging timestamps that deviate from these patterns as anomalies. However, these approaches often fail to capture salient intra-variate temporal and inter-variate dependencies in time series due to their focus on excessively fine granularity, leading to suboptimal performance. In this study, we introduce MtsCID, a novel semi-supervised multivariate time series anomaly detection method. MtsCID employs a dual network architecture: one network operates on the attention maps of multi-scale intra-variate patches for coarse-grained temporal dependency learning, while the other works on variates to capture coarse-grained inter-variate relationships through convolution and interaction with sinusoidal prototypes. This design enhances the ability to capture the patterns from both intra-variate temporal dependencies and inter-variate relationships, resulting in improved performance. Extensive experiments across seven widely used datasets demonstrate that MtsCID achieves performance comparable or superior to state-of-the-art benchmark methods. Our code is available at https://github.com/ilwoof/MtsCID/ . DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation emai; April 2025; Sydney, Australia ISBN: 978-1-4503-XXXX-X/18/06 CCS: Software and its engineering Software maintenance tools CCS: Computer systems organization Reliability CCS: Computer systems organization Availability CCS: Computer systems organization Maintainability and maintenance 1. Introduction Modern society increasingly relies on web-based application systems integrated into distributed systems, cloud computing platforms, and Internet of Things (IoT) devices ( Luo et al., 2021 ; Wang et al., 2024 ; Nam et al., 2024 ) . These systems function across various sectors, including finance, transportation, telecommunications, media, and industrial operations. Downtime or service interruptions in these critical infrastructures can disrupt daily life, create chaos in business operations, and result in significant financial losses ( Elliot, 2014 ; Ma et al., 2020 ; Zhang et al., 2024 ) . To ensure high reliability and availability of these systems, numerous AI-based methods ( Breunig et al., 2000 ; Schölkopf et al., 2001 ; Liu et al., 2008 ; Zong et al., 2018 ; Su et al., 2019 ; Audibert et al., 2020 ; Shen et al., 2020 ; Xu, 2021 ; Tuli et al., 2022 ; Chen et al., 2022 ; Liu et al., 2023 ; Lee et al., 2023 ; Xiao et al., 2023 ; Li et al., 2023 ; Yang et al., 2023 ; Nam et al., 2024 ; Chen et al., 2024 ) have been proposed to detect anomalies from system operational data, such as service KPIs, as well as system runtime status data like CPU and memory usage. This data often takes the form of Multivariate Time Series (MTS). MTS anomaly detection aims to identify whether time steps in a series are normal or abnormal. Despite the vast amount of time series generated daily in these systems, supervised learning methods often face challenges in this domain due to the labor-intensive data labeling process and the scarcity of anomalous instances ( Xu, 2021 ; Li et al., 2023 ) . As a result, MTS anomaly detection is typically framed as a semi-supervised learning task, where only normal data is available for model training. Traditional machine learning methods, such as Local Outlier Factor (LOF) ( Breunig et al., 2000 ) , One Class Support Vector Machine (OCSVM) ( Schölkopf et al., 2001 ) , Isolation Forest (iForest) ( Liu et al., 2008 ) , have been widely used for anomaly detection tasks. These methods treat multivariate observations at time steps as points in a feature space. Distance or density metrics are used to assess the proximity of these points to each other. The points that deviate significantly from the majority are flagged as anomalies. We refer to this approach as proximity-based. Recently, proximity-based methods that combine deep representation learning, such as DeepSVDD ( Ruff et al., 2018 ) and DAGMM ( Zong et al., 2018 ) , have also emerged. However, this approach often struggles with high accuracy due to their inability to effectively capture dynamic intra-variate temporal dependencies and complex inter-variate relationships. To address the aforementioned issues, many temporal-based and spatiotemporal-based methods have been developed ( Shen et al., 2020 ; Xu, 2021 ; Li et al., 2021 ; Li et al., 2023 ; Yang et al., 2023 ; Song et al., 2024 ; Chen et al., 2024 ) . For instance, THOC ( Shen et al., 2020 ) employs a differentiable hierarchical clustering mechanism to integrate temporal features across various scales and resolutions for effective normal pattern learning. AT ( Xu, 2021 ) models prior and series associations between time steps in series for temporal pattern learning. DCdetector ( Yang et al., 2023 ) utilizes two patch-based attention networks for contrastive learning to capture temporal dependencies in the given time series. We classify these approaches as temporal-based methods since they primarily rely on temporal dependencies for anomaly detection. In contrast, InterFusion ( Li et al., 2021 ) leverages a hierarchical Variational Autoencoder framework to capture both intra-variate temporal and inter-variate dependencies. Memto ( Song et al., 2024 ) presents a memory-guided reconstruction approach that utilizes a single Transformer network to capture temporal dependencies, while also integrating inter-variate associations through the interactions between the derived representations and a set of memory items. STEN ( Chen et al., 2024 ) presents a framework that combines subsequence order prediction, capturing temporal correlations, with distance prediction, which learns spatial relationships between sequences. These methods are categorized into spatiotemporal-based methods. While both temporal-based and spatiotemporal-based methods enhance the ability to capture intricate intra-variate or/and inter-variate dependencies within time series, they often fail to capture salient intra-variate temporal and inter-variate dependencies in time series due to their focus on excessively fine granularity. This limitation can result in their suboptimal performance in MTS anomaly detection. This paper presents MtsCID, a novel semi-supervised anomaly detection approach for M ultivariate T ime S eries through capturing Coarse -grained I ntra-variate and inter-variate D ependencies. MtsCID employs a dual-network architecture: one network utilizes the attention maps of multi-scale intra-variate patches to capture coarse-grained temporal dependencies between time steps, while the other focuses on variate interactions, leveraging convolutions, frequency component-based Transformer and a set of sinusoidal prototypes to capture coarse-grained inter-variate relationships. The deviation from normal patterns in each dimension is aggregated to generate losses during training and anomaly scores for each time step during inference. The resulting anomaly score indicates whether a given timestamp is anomalous or not. Our approach has been evaluated on seven commonly used publicly available datasets. Experimental results demonstrate its effectiveness, achieving comparable or superior anomaly detection performance to nine state-of-the-art methods. In summary, our main contributions are as follows: (1) A novel time-frequency interleaved learning scheme: We introduce an novel scheme for learning intra- and inter-variate dependencies through interleaved processing in both

Read the complete text