Credibility-Weighted Graph Analytics for Behavioral Intention Forecasting Under Disrupted Information Ecosystems
Main article
Abstract
Behavioral intention forecasting in networked social environments has grown substantially more difficult as digital ecosystems become subject to sudden credibility disruptions such as data breaches, institutional scandals, and coordinated misinformation campaigns. Conventional graph learning methods assign homogeneous influence weights to network edges and therefore struggle to capture how the persuasive capacity of a social tie degrades or recovers when the source actor loses perceived legitimacy. This paper proposes a credibility-weighted graph analytics (CWGA) framework that explicitly represents node-level and edge-level credibility as dynamic, learnable quantities that modulate attention during message passing on temporal interaction graphs. The framework introduces a credibility propagation layer that updates source scores at each observed event using content quality, sentiment polarity, and institutional role signals, and a disruption-aware gating mechanism that suppresses or amplifies pathway contributions in response to detected shock events. Intention is subsequently predicted from the joint representation of interaction history and time-varying credibility state. Experimental evaluation on four large-scale synthetic network scenarios with calibrated disruption regimes demonstrates that CWGA improves AUC by 2.1 to 5.8 percentage points over the strongest temporal graph baselines in stable conditions and by 3.4 to 8.6 percentage points under active disruption, while regression error decreases by 11 to 22 percent on the same scenarios. An ablation study confirms that credibility propagation and disruption-aware gating each contribute independently to performance, with their combination yielding the largest gains specifically in community-structured networks where bridge actors serve as cross-module credibility conduits. These findings establish credibility dynamics as a first-class feature for behavioral analytics and provide a reproducible framework for intention modeling under realistic information ecosystem conditions.
