Chi-Square Wavelet Graph Neural Networks for Heterogeneous Graph Anomaly Detection
Xiping Li, Xiangyu Dong, Xingyi Zhang, Kun Xie, Yuanhao Feng, Bo Wang, Guilin Li, Wuxiong Zeng, Xiujun Shu, Sibo Wang · 2025
Read Chi-Square Wavelet Graph Neural Networks for Heterogeneous Graph Anomaly Detection by Xiping Li, Xiangyu Dong, Xingyi Zhang, Kun Xie, Yuanhao Feng, Bo Wang, Guilin Li, Wuxiong Zeng, Xiujun Shu, Sibo Wang (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-2505-18934.txt.
Introduction
Xiping Li and co-authors present ChiGAD, a graph neural network that spots anomalies in mixed-type networks. Real networks mix many node and edge types, which breaks methods built for uniform graphs. Three gaps stand out: weak capture of odd signals, loss of sharp detail, and poor learning from rare cases.
ChiGAD uses a chi-square filter drawn from wavelet ideas in signal work. One part applies a tuned filter to each meta-path graph. A second part aligns features while keeping high-frequency detail. A third part weights the loss toward hard anomaly cases. Together the parts aim to catch rare fraud or faults in complex data. The design suits security and finance graphs.
Catalog record
| CREATOR | Xiping Li, Xiangyu Dong, Xingyi Zhang, Kun Xie, Yuanhao Feng, Bo Wang, Guilin Li, Wuxiong Zeng, Xiujun Shu, Sibo Wang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-18934
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Chi-Square Wavelet Graph Neural Networks for Heterogeneous Graph Anomaly Detection Xiping Li, Xiangyu Dong, Xingyi Zhang, Kun Xie, Yuanhao Feng, Bo Wang, Guilin Li, Wuxiong Zeng, Xiujun Shu, Sibo Wang Abstract . Graph Anomaly Detection (GAD) in heterogeneous networks presents unique challenges due to node and edge heterogeneity. Existing Graph Neural Network (GNN) methods primarily focus on homogeneous GAD and thus fail to address three key issues: (C1) Capturing abnormal signal and rich semantics across diverse meta-paths; (C2) Retaining high-frequency content in HIN dimension alignment; and (C3) Learning effectively from difficult anomaly samples with class imbalance. To overcome these, we propose ChiGAD, a spectral GNN framework based on a novel Chi-Square filter, inspired by the wavelet effectiveness in diverse domains. Specifically, ChiGAD consists of: (1) Multi-Graph Chi-Square Filter, which captures anomalous information via applying dedicated Chi-Square filters to each meta-path graph; (2) Interactive Meta-Graph Convolution, which aligns features while preserving high-frequency information and incorporates heterogeneous messages by a unified Chi-Square Filter; and (3) Contribution-Informed Cross-Entropy Loss, which prioritizes difficult anomalies to address class imbalance. Extensive experiments on public and industrial datasets show that ChiGAD outperforms state-of-the-art models on multiple metrics. Additionally, its homogeneous variant, ChiGNN, excels on seven GAD datasets, validating the effectiveness of Chi-Square filters. Our code is available at https://github.com/HsipingLi/ChiGAD. 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: 10.1145/3711896.3736877 ISBN: 979-8-4007-1454-2/2025/08 CCS: Computing methodologies Anomaly detection † † cc-license: by 1. Introduction Graph Anomaly Detection (GAD) ( Tang et al., 2023 ) has emerged as a critical research focus of graph learning , with vital applications ranging from identifying risky merchants ( Liu et al., 2022 ) to detecting fake news, ( Shu et al., 2017 ; Zhang et al., 2025 ) , and malicious users ( Liu et al., 2018 ; Wang et al., 2017 ) . Graph Neural Networks (GNNs) ( Kipf and Welling, 2017 ; Velickovic et al., 2018 ) have become the preferred solution for GAD tasks ( Tang et al., 2022 ; Gao et al., 2023 ) , given their inherent ability to process graph-structured data. Since node-level GAD is fundamentally a binary node classification task, thorough assessments of various traditional GNN models ( Kipf and Welling, 2017 ; Hamilton et al., 2017 ) have been conducted. Recent spectral GNNs ( Tang et al., 2022 ; Gao et al., 2023 ; Chai et al., 2022 ) , especially those employing wavelet filters ( Tang et al., 2022 ; Lu and Ghorbani, 2009 ; Alarcon-Aquino and Barria, 2001 ) , have shown great superiority in homogeneous GAD tasks. Yet, these approaches invariably assume unified node and edge types of fraud graphs, ignoring the scenarios characterized by Heterogeneous Information Networks (HINs). Despite the successful application of GNN-based models in various HINs, including social network dynamics ( Zhang et al., 2015 ; Kong et al., 2013 ) , link prediction ( Li et al., 2018 ; Davis et al., 2011 ) , and recommendations ( Dong et al., 2012 ; Cantador et al., 2011 ; Li et al., 2024 ; Yu et al., 2014 ; Shi et al., 2019 ) , anomaly detection in heterogeneous networks remains understudied. Considering the significance of GAD in critical HINs such as financial networks, a specialized framework for heterogeneous GAD is urgently needed. Learning on HINs is inherently challenging due to both the node and edge heterogeneity ( Zhang et al., 2019 ) . Recently, a series of meta-path-based GNNs ( Wang et al., 2019 ; Fu et al., 2020 ; Zhu et al., 2019 ; Zhang et al., 2019 ; Hu et al., 2020 ) and multi-graph filters ( He et al., 2024 ; Butler et al., 2023 ) have been proposed to address the issues arising from heterogeneity. However, it is worth noting that they overlook the following critical challenges due to not being specifically designed for GAD tasks. Particularly, (C1) Capturing abnormal signals and rich semantics across diverse meta-paths: In GAD applications such as fraud detection, meta-paths exhibit distinct frequency characteristics crucial for anomaly identification. For instance, in a financial network, transaction-based meta-paths may show irregular bursts, while social interactions remain steady. A generic filter cannot capture such differences across frequency bands. Previous works ( Butler et al., 2023 ; He et al., 2024 ) adopt a unified signal processing framework, thereby overlooking the nuanced frequency energy distributions inherent in different meta-path graphs. (C2) Retaining high-frequency content in HIN dimension alignment: The retention of high-frequency content is essential to avert over-smoothing and performance degradation in homogeneous GAD tasks ( Tang et al., 2022 ; Gao et al., 2023 ) ; additionally, the node heterogeneity inherent in HIN engenders dimensional disparity, compelling nodes to depend on a prior dimensional alignment technique for feasible representation learning within heterogeneous graphs ( Zhang et al., 2019 ) . Nevertheless, the risky potential for the alignment method to induce loss of high-frequency content within these signals remains an unexplored domain since they directly process graph signals. (C3) Learning effectively from difficult anomaly samples with class imbalance: In GAD tasks, the inherent class imbalance poses significant challenges. This imbalance makes it difficult for models to learn subtle anomaly characteristics. While existing studies ( Dong et al., 2024 ) have acknowledged this challenge, a comprehensive strategy addressing class imbalance while effectively learning from rare and difficult anomaly samples is still lacking. Inspired by the demonstrated effectiveness of wavelet in various domains ( Monti et al., 2017 ; Xu et al., 2019a ) , especially in homogeneous GAD ( Tang et al., 2022 ; Lu and Ghorbani, 2009 ; Alarcon-Aquino and Barria, 2001 ) , we first employ the Chi-Square distribution to design a novel graph filter, namely, Chi-Square filter . This novel graph wavelet filter not only inherits the proven wavelet properties crucial for GAD tasks ( Tang et al., 2022 ; Lu and Ghorbani, 2009 ; Alarcon-Aquino and Barria, 2001 ) , but also possesses a unique property of additivity compared to other wavelets, which makes Chi-Square naturally suitable for collaborative graph representation learning across diverse meta-paths within HIN, as elaborated in Section 4.1 . Furthermore, motivated by the aforementioned three crucial challenges, we introduce a novel framework for heterogeneous graph anomaly detection, namely ChiGAD , which takes our proposed Chi-Square filter as a basic component and is comprised of three essential components: (1) Multi-Graph Chi-Square Filter : It processes different node types independently. For each type of nodes, it leverages the additivity of Chi-Square to collaboratively capture fine-grained semantics across diverse meta-paths by assigning the most aligned Chi-Square filters according to the spectral energy distributions of representative meta-path graphs; (2) Interactive Meta-Graph Convolution : It addresses representation dimension disparities in heterogeneous graphs by using a linear layer to align feature dimensions, justified by its ability to preserve high-frequency components to resist over-smoothing, which is helpful to heterogeneous GAD tasks, as proven in Section 4.3 in theory and analyzed in Appendix 5.4 . It aligns node representations across types and employs a unified Chi-Square filter on a