LLMBot: Dual-View LLM-Enhanced Social Bot Detection with Structure-Conditioned Prompting and Cross-View Alignment

摘要

Social bot detection is increasingly a relational trust problem: bots write fluent text and attach themselves to genuine communities, so the observed graph contains many incidental and adversarially injected edges. Graph detectors that treat every edge as informative aggregate misleading neighbors, while heterophily-oriented or structurally robust GNNs decide which edges to trust from topology alone—they cannot explain why a specific tie is suspicious. We propose LLMBot, which transforms an LLM's account-level semantic judgment into an edge-level reliability signal that gates relational message passing. Conditioned on a compact structural interface rather than verbose neighborhood serialization, the LLM produces bounded structured artifacts for each account and sampled tie. A lightweight interpreter distills these artifacts into cacheable embeddings, while tie-level reliability scores control gated neighborhood aggregation. A cross-view consistency objective, adapted from established alignment techniques, prevents either the semantic or structural view from dominating when one information channel is corrupted. Under the official, naturally class-imbalanced TwiBot-22 protocol, LLMBot achieves a bot-F1 score of 80.82, outperforming the strongest non-LLM graph baseline by +16.58 points. On the smaller, near-balanced TwiBot-20 benchmark, LLMBot performs on par with the state of the art, achieving 89.30 versus 89.10, with the difference falling within seed variability, while also providing better calibration. Semantic gating outperforms structural gating in all 44 evaluated perturbation settings, covering random edge noise and three targeted attacks, with FAGCN, GPRGNN, DropEdge, and GNNGuard independently substituted into the same framework. The improvement exceeds the pooled standard deviation across seeds in 39 settings. The remaining five statistically comparable settings occur only on the easier benchmark, where the proposed mechanism predicts limited additional benefit. The learned reliability weights distinguish injected edges from genuine edges with AUROC scores of 0.846 and 0.872 on TwiBot-20 and TwiBot-22, respectively. All LLMBot experiments use identity-anonymized inputs. In strictly paired experiments, restoring account identifiers improves bot-F1 by only +2.00 points, bounding the extent to which memorization of public accounts by the LLM could explain the observed gains. Code and data are available at: https://anonymous.4open.science/r/LLMBot.