Multi-Order Hyperbolic Graph Convolution and Aggregated Attention for Social Event Detection

TL;DR

The paper introduces MOHGCAA, a novel framework that integrates multi-order graph convolution with aggregated attention in hyperbolic space to address social event detection. By projecting Euclidean features into hyperbolic space, performing multi-order convolutions in the tangent space, and aggregating via attention before mapping back, the approach captures hierarchical and higher-order relationships that Euclidean methods miss. The authors present both unsupervised and supervised variants, validate on multiple datasets, and show consistent improvements over strong baselines, with hyperbolic space and model variants yielding nuanced gains. The work advances hyperbolic representation learning for SED, offering a scalable, robust method that improves detection in tree-like, hierarchical social data and suggesting directions for real-time and broader graph-based tasks.

Abstract

Social event detection (SED) is a task focused on identifying specific real-world events and has broad applications across various domains. It is integral to many mobile applications with social features, including major platforms like Twitter, Weibo, and Facebook. By enabling the analysis of social events, SED provides valuable insights for businesses to understand consumer preferences and supports public services in handling emergencies and disaster management. Due to the hierarchical structure of event detection data, traditional approaches in Euclidean space often fall short in capturing the complexity of such relationships. While existing methods in both Euclidean and hyperbolic spaces have shown promising results, they tend to overlook multi-order relationships between events. To address these limitations, this paper introduces a novel framework, Multi-Order Hyperbolic Graph Convolution with Aggregated Attention (MOHGCAA), designed to enhance the performance of SED. Experimental results demonstrate significant improvements under both supervised and unsupervised settings. To further validate the effectiveness and robustness of the proposed framework, we conducted extensive evaluations across multiple datasets, confirming its superiority in tackling common challenges in social event detection.

Figures (11)

• Figure 1: (a) The points A and B in Euclidean space, situated on distinct branches of the diagram, are in greater proximity, but points A’ and B’ in hyperbolic space, also on different branches, exhibit a more reasonable closeness to one another. (b) A diagram illustrating the characteristics of hyperbolic space and its tangent plane, where lines e1 and e2 are two parallel lines passing through point c, which do not conform to the fifth axiom of contemporary geometry.
• Figure 2: Depictions of three hyperbolic graphical models: the Lorentz model, the Kälin model, and the Poincaré ball model.
• Figure 3: The $\exp_o(\cdot)$ operation maps data from Euclidean space to hyperbolic space, whereas the $\log_o(\cdot)$ operation maps data from hyperbolic space to its Euclidean tangent plane.
• Figure 4: Upon initialization of the data phase in hyperbolic space, it is projected into the tangent plane of its o-points via the $\exp_o(\cdot)$ function, and subsequent to the aggregation manipulate, it is implicitly re-mapped into the new hyperbolic space using the $\log_o(\cdot)$ function.
• Figure 5: The overall framework of this study: First, the data representation $\mathrm{X} _i$ and its corresponding adjacency matrix are obtained. Next, $\mathrm{X} _i$ is mapped into hyperbolic space via the $\exp_o(\cdot)$ function, producing its hyperbolic representation. Then, the $\log_o(\cdot)$ function is applied to project $\mathrm{X} _i$ onto its tangent space, yielding the representation $x _i$.On this basis, a multi-order graph convolution network is employed to derive the multi-order representation $h_i^k$. Subsequently, an attention-based network is used to generate the hyperbolic multi-order graph representation $h_i$. Finally, the convolution attention representation $h_i$ is mapped into a new hyperbolic space through the $\log_o(\cdot)$ function, resulting in the hyperbolic representation $\mathcal{H} _i$.