Title: SAHG: A Sector-Anisotropic Hyperbolic Graph Approach to Social Bot Identification

Abstract:

The rise of large language model (LLM)-powered social bots has generated fluent, human-mimicking text, thereby diminishing the effectiveness of detection methods that rely solely on content analysis. Nevertheless, coordinated bot campaigns inevitably produce distinct relational patterns—such as shared neighborhoods, behavioral similarities, specific community positions, coordinated activities, and interaction networks—that graph-based detection strategies can leverage. However, current graph-based detectors encounter two primary obstacles when utilizing such evidence.

First, conventional Euclidean Graph Neural Networks (GNNs) fail to accurately represent the hierarchical and scale-free nature of social graphs. Although hyperbolic geometry mitigates issues related to volume growth mismatches, models with fixed curvature assign uniform geometric resolution to all structural directions, ignoring differences in density and separation requirements. Second, relational data is not always trustworthy. Advanced bots often establish heterophilic links with authentic users, which causes neighborhood aggregation to blend bot and human signals, ultimately diluting account-level evidence.

To overcome these limitations, we introduce SAHG (Sector-Anisotropic Hyperbolic Graph). SAHG addresses both challenges by learning a direction-dependent curvature field, $\gamma(u)$, which adapts geometric resolution across various structural directions. It further employs sector prototypes to transform angular concentration and alignment into features suitable for classification. To ensure that contaminated neighborhood aggregation does not overpower account-level signals, SAHG encodes per-account features and graph-neighborhood representations within two separate SAH channels, merging them only at the final classifier stage.

Evaluations on the Fox8-23, BotSim-24, and MGTAB datasets demonstrate that SAHG attains the highest accuracy and F1 scores among all tested benchmarks. It surpasses a range of baselines, including feature-based, graph-based, LLM-based, and isotropic hyperbolic models. Furthermore, ablation studies and geometric analyses validate the efficacy of both the anisotropic geometry and the dual-channel architecture.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC