HLSAD: Hodge Laplacian-based Simplicial Anomaly Detection

Florian Frantzen, Michael T. Schaub · 2025

Read HLSAD: Hodge Laplacian-based Simplicial Anomaly Detection by Florian Frantzen, Michael T. Schaub (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-2505-24534.txt.

Introduction

Florian Frantzen and Michael Schaub present HLSAD, a method for spotting anomalies in networks whose links involve more than two points at once. Their paper appeared in 2025. Ordinary graph tools flatten such multi-way links and miss structural changes that live in triangles and larger groups. The new method reads the spectrum of Hodge Laplacians, operators built for these higher-order structures, to model how signals sit across them through time.

Tested on synthetic and real data, HLSAD beat existing graph methods at catching both sudden events and change points. The authors say the spectral view is also computationally efficient. The work targets settings like traffic, biology, and finance, where group interactions, not just pairs, carry the story.

Catalog record

CREATORFlorian Frantzen, Michael T. Schaub
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-24534

Browse the full library · About the Tower of Babel

Full text

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

HLSAD: Hodge Laplacian-based Simplicial Anomaly Detection Florian Frantzen, Michael T. Schaub Abstract . In this paper, we propose HLSAD, a novel method for detecting anomalies in time-evolving simplicial complexes. While traditional graph anomaly detection techniques have been extensively studied, they often fail to capture changes in higher-order interactions that are crucial for identifying complex structural anomalies. These higher-order interactions can arise either directly from the underlying data itself or through graph lifting techniques. Our approach leverages the spectral properties of Hodge Laplacians of simplicial complexes to effectively model multi-way interactions among data points. By incorporating higher-dimensional simplicial structures into our method, our method enhances both detection accuracy and computational efficiency. Through comprehensive experiments on both synthetic and real-world datasets, we demonstrate that our approach outperforms existing graph methods in detecting both events and change points. Conference: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 3–7, 2025; Toronto, ON, Canada Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’25), August 3–7, 2025, Toronto, ON, Canada DOI: 10.1145/3711896.3736998 ISBN: 979-8-4007-1454-2/2025/08 CCS: Mathematics of computing Graph theory CCS: Computing methodologies Anomaly detection CCS: Computing methodologies Temporal reasoning CCS: Computing methodologies Spectral methods CCS: Theory of computation Dynamic graph algorithms CCS: Mathematics of computing Topology Teaser figure illustrating HLSAD on a toy example. The figure shows a flowchart of the process: First, the input is a sequence of simplicial complexes, which are then processed to compute the Hodge Laplacians. Next, we extract spectral features from the Hodge Laplacians and collect them into context matrices for each timestep. These are processed into typical spectral features and an anomaly score is computed for each timestep by comparing the current spectral features to a characteristic feature vector. Finally, anomalies are detected based on the computed scores. Figure 1. Illustrative process of anomaly detection on a toy-example simplicial complex sequence. Starting from a time-evolving simplicial complex, we compute the Hodge Laplacians for each snapshot and extract spectral features for each. We employ a sliding window approach to identify anomalies by evaluating the deviation of current spectral features from established temporal patterns. In this illustrative example, HLSAD successfully detects all three artificially introduced anomalies, whereas traditional graph-based analysis fails to identify the final anomaly that manifests in higher-order interactions. Teaser figure illustrating HLSAD on a toy example. The figure shows a flowchart of the process: First, the input is a sequence of simplicial complexes, which are then processed to compute the Hodge Laplacians. Next, we extract spectral features from the Hodge Laplacians and collect them into context matrices for each timestep. These are processed into typical spectral features and an anomaly score is computed for each timestep by comparing the current spectral features to a characteristic feature vector. Finally, anomalies are detected based on the computed scores. † † cc-license: by KDD Availability Link: The source code of this paper has been made publicly available at https://doi.org/10.5281/zenodo.15511585 and at https://git.rwth-aachen.de/netsci/hlsad . 1. Introduction Anomaly detection within time-evolving networks has widespread applications across diverse domains, including network security for detecting intrusions and system faults ( Idé and Kashima, 2004 ; Sun et al., 2008 ; Ding et al., 2012 ) , social network analysis for identifying collective behavioral changes induced by external factors ( Peel and Clauset, 2015 ; Yu et al., 2016 ) , and ecological monitoring for detecting ecosystem disturbances ( Cheng et al., 2008 ) . Accordingly, many graph-based techniques have been developed for identifying events and change points in dynamic networks ( Akoglu et al., 2015 ; Ranshous et al., 2015 ; Zhou et al., 2024 ) . Despite their utility, graph-based approaches are however limited by their focus on dyadic relationships, and fail to capture higher-order interactions that can be essential for identifying complex structural anomalies in the underlying system. Such higher-order data can manifest itself through different mechanisms: Higher-Order Data: Many real-world datasets inherently contain multi-way interactions that extend beyond dyadic relationships. Traditional graph-based methods are fundamentally limited in their ability to capture and analyze these higher-order structures. For instance, during the COVID-19 pandemic, social distancing measures disproportionately affected large group gatherings compared to pairwise interactions, creating structural changes that are more readily detectable through higher-order methods ( Pouw et al., 2020 ) . Lifting: Even when data only contains pairwise relations between entities, we can employ graph lifting techniques to construct a simplicial complex representation. A common approach identifies cliques in the graph with equivalent simplices. The effectiveness of lifting has been demonstrated in various contexts; notably, the Weisfeiler-Lehman isomorphism test exhibits strictly greater discriminative power when applied to clique complexes compared to their graph counterparts ( Bodnar et al., 2021 ) . Building upon the graph anomaly detection frameworks developed by Akoglu and Faloutsos (2010) and Huang et al. (2020) , we propose a novel method that leverages simplicial complexes to capture and analyze multi-way interactions in temporal data sequences. It is important to distinguish our focus on temporal anomaly detection in time-evolving networks from the related but distinct problem of identifying anomalous substructures within static networks, as these two problems address fundamentally different analytical objectives despite their similar names. The incorporation of higher-order interactions via simplices in our method provides significant advantages over traditional graph-based approaches: First, it enhances detection accuracy through the explicit modeling of higher-order interactions that are inherently absent in graph representations. These higher-order structures are particularly relevant in domains such as epidemic modeling, where large group gatherings may exhibit distinct response patterns to intervention measures compared to dyadic interactions. Second, even if we only have pairwise interaction data, the application of graph lifting techniques results in improved anomaly detection capabilities. Our experimental results indicate that our method achieves superior accuracy when applied to lifted graph datasets compared to conventional spectral approaches on graph skeletons and can even offer enhanced computational efficiency. HLSAD operates through the following systematic process: For each snapshot within a temporal sequence of simplicial complexes, we compute the corresponding Hodge Laplacians up to a user-specified order. Subsequently, we extract the principal eigenvalues from both up and down Hodge Laplacians and concatenate these spectral features into a comprehensive feature vector for each temporal instance. Employing a sliding window mechanism, we derive a characteristic feature vector by computing a weighted average across the window’s temporal span. When a new feature vector exhibits significant deviation from this established characteristic pattern, we designate the corresponding temporal instance as an anomaly. We illustrate this systematic process in Figure 1 and highlight a case where our higher-order approach successfully identifies an anomaly that is overlooked by traditional graph-based methods. Contribution To the best of our knowledge, this paper presents the first methodology for anomaly detection in temporal sequences of simplicial complexes. Our experimental results demonstrate that leveraging simplicial complex representations yields superior performance compared to traditional graph-based approaches in two key scenarios: First, when analyzing datasets containing inherent higher-order interactions, and second, when applying graph lifting techniques to enrich the topological structure. Outline The remainder of this paper is organized as follows: In Section 2 , we review the existing literature on graph anomaly detection methods. In Section 3 , we introduce the mathematical framework and

Read the complete text