Abstract
Detecting abnormal events in complex networks is a critical task in fields ranging from academic collaboration to e-commerce. Traditional methods often focus on pair-wise interactions, failing to capture the "set-wise" complexity of real-world events. This paper presents AEHCL, a framework that models events as hypergraphs and utilizes contrastive learning to identify anomalies. By analyzing intra-event interactions and inter-event neighborhood semantics, AEHCL effectively identifies events that may appear normal in isolation but are anomalous in their specific combination of participants or context.
Introduction
In modern data science, many real-world scenarios are best represented as Attributed Heterogeneous Information Networks (AHINs). These networks involve multiple types of entities—such as authors, papers, and venues—each possessing rich attributes. A significant challenge in these networks is detecting "abnormal events."
Most existing anomaly detection techniques focus on pair-wise relationships (e.g., "Is it unusual for Person A to interact with Person B?"). However, anomalies are often multivariate. For instance, three world-class data scientists are individually "normal" entities, but if they suddenly co-author a paper on radiology for a bio-medicine conference, the combination of these entities constitutes an anomaly. To capture these set-wise discrepancies, we must model events as whole units—hyperedges—rather than a collection of independent pairs.
Methodology: The AEHCL Framework
AEHCL (Abnormal Event detection via Hypergraph Contrastive Learning) is an unsupervised framework designed to detect these complex anomalies without the need for labeled data. It operates through three distinct contrastive modules that evaluate an event from different perspectives:
1. Intra-event Pair-wise Module
This module serves as the foundational check, ensuring that the individual pairs within an event are compatible. It learns to distinguish between "normal" internal pairings and random, mismatched pairs, ensuring that the basic building blocks of the event are semantically sound.
2. Intra-event Multivariate Module
This is the core "set-wise" detector. The framework identifies a center node (e.g., a Paper) and its associated context nodes (e.g., Authors and Venue). By employing self-attention mechanisms and max-pooling, the module calculates a compatibility score for the entire group. If the collective "fit" of the group is poor, the compatibility score drops, flagging a potential anomaly.
3. Inter-event Module
An event might be internally consistent but still be a "fish out of water" within its broader environment. This module examines the "neighborhood" of events using meta-paths (predefined relational patterns). It compares a target event to semantically similar neighbors; an event that stands out as a local mismatch is flagged as an anomaly.
Finally, AEHCL aggregates these signals into a specialized anomaly score function, prioritizing the most significant mismatches to generate a final ranking of suspicious events.
Experimental Setup
The model was evaluated against three real-world AHIN datasets:
- Aminer: An academic network (Center: Paper; Context: Authors, Venues).
- IMDB: A movie network (Center: Movie; Context: Directors, Actors).
- Meituan: A food delivery network (Center: Order; Context: Shops, Users).
Because natural anomaly labels are rare, the authors utilized an Abnormal Event Injection strategy. This involved replacing nodes in legitimate events with "distant" nodes—selected based on Euclidean distance of attributes—to create challenging and realistic synthetic anomalies for testing.
Results and Discussion
AEHCL demonstrated superior performance compared to state-of-the-art baselines, including specialized models like CoLA and ANEMONE. Key performance metrics include:
- Average Precision (AP): Improved by up to 12.0%.
- Area Under Curve (AUC): Improved by up to 4.6%.
A case study on the Aminer dataset highlighted the model's practical utility. AEHCL successfully identified a paper where data mining specialists published in a bio-medicine venue. While the authors and the venue were individually reputable, their combination was rare, representing a classic multivariate anomaly that pair-wise models often miss.
Limitations and Future Directions
While effective, the model's performance on the Aminer dataset showed sensitivity to negative sampling. When the number of negative samples exceeded 80, performance dipped, suggesting that random sampling might occasionally select "negative" instances that are too similar to the target, creating noise. Additionally, like most anomaly detection systems, selecting the optimal threshold to distinguish between normal and abnormal events remains a task that requires domain-specific tuning.
Future work will focus on developing more sophisticated negative sampling strategies to further enhance the model's discriminative power and robustness across diverse network types.
Conclusion
By shifting the focus from pair-wise interactions to hypergraph-based set-wise analysis, AEHCL provides a more nuanced and accurate approach to anomaly detection in heterogeneous networks. Its ability to integrate intra-event compatibility with inter-event semantic context makes it a powerful tool for identifying complex, hidden anomalies in large-scale datasets.
Summary updated