TLDR: Janus is a novel framework for node-level anomaly detection that integrates Euclidean and Hyperbolic Graph Neural Networks. It processes nodes using two distinct views (original and structural features) and embeds them into both geometric spaces. A multi-Graph-Autoencoder, enhanced with a contrastive learning objective, then identifies anomalies as nodes whose representations are challenging to reconcile across these diverse geometric perspectives. Experimental results on four real-world datasets demonstrate that Janus consistently surpasses existing state-of-the-art methods.
In the evolving landscape of machine learning, identifying unusual patterns or “anomalies” within complex datasets is crucial for many real-world applications. Think of detecting fraudulent transactions, pinpointing cyber threats, or even refining recommendation systems. When this data is structured as a graph – where individual items (nodes) are connected by relationships (edges) – the challenge becomes even more intricate. This is known as Node-level Anomaly Detection (NAD).
A new research paper introduces an innovative framework called Janus, designed to significantly improve how we spot these anomalies in graphs. The core idea behind Janus is to combine two distinct geometric perspectives: Euclidean and Hyperbolic representations. While Euclidean geometry is what we’re familiar with in our everyday world (flat spaces, straight lines), Hyperbolic geometry deals with spaces that have a constant negative curvature, making it particularly good at representing hierarchical structures and complex relationships that don’t fit neatly into a flat space.
Janus works by creating two “views” for each node in a graph. One view uses the node’s original features, while the other captures its structural characteristics derived from things like random walks and node degrees. These two views are then embedded into both Euclidean and Hyperbolic spaces using specialized Graph Neural Networks (GNNs). The framework employs a multi-Graph-Autoencoder, which is essentially a system that learns to encode and decode these node representations. Crucially, it’s equipped with a contrastive learning objective. This objective helps align the embeddings across the Euclidean and Hyperbolic spaces. Nodes whose views are difficult to reconcile – meaning they don’t align well across these different geometric interpretations – are flagged as potentially anomalous.
The researchers conducted extensive experiments on four real-world datasets, including Disney, Books, Reddit, and T-Finance, which feature naturally occurring anomalies rather than artificially generated ones. The results were compelling: Janus consistently outperformed both traditional “shallow” methods and more recent deep learning baselines. This empirical evidence strongly suggests that leveraging multiple geometric representations provides a more robust and effective way to identify subtle and complex anomalies within graphs.
Also Read:
- Unraveling Anomalies: A New Approach to Causal Disentanglement in Time Series Data
- Enhancing AI Reliability: A New Method for Detecting Unexpected Data
This breakthrough in combining different geometric spaces for anomaly detection opens new avenues for understanding and tackling complex data irregularities. For those interested in the technical details, the full research paper is available here: Combining Euclidean and Hyperbolic Representations for Node-level Anomaly Detection.
