Uncertainty Calibrated Cross Platform Identity and Dyadic Episode Modeling for Cyberbullying Analysis

Abstract

Cyberbullying is an important issue in digital safety because hostile communications remain accessible on the Internet, intensify over time, and can have profound psychological and social impacts. Timely and proportional moderation is therefore crucial and cannot be achieved without accurate automated detection. Existing studies rely on attention networks, recurrent networks, contextual embeddings, temporal fusion, graph networks, and metaheuristic optimization, but the majority take only single-stage text content into consideration and fail to collectively verify the repeated targeting behavior, cross-platform identity consistency, dynamic power imbalance, harm-directedness, and uncertainty of prediction. To fill this gap, this study presents Causal Identity-aware Dyadic Episode Reasoning for Cyberbullying (CIDER-CB), an advanced learning framework that contains three phases. Phase I conducts multimodal representation with reliability-gating, temporal heterogeneous-graph construction, and probabilistic cross-platform identity alignment. Phase II deploys a marked temporal point process to measure repetition and escalation, a graph-based dyadic reasoning approach to measure the power asymmetry between perpetrator and victim dyad components, and counterfactual causal learning to measure harm towards a target. Phase III combines fairness-constrained intervention learning, evidential uncertainty estimation, and concept drift-aware continual adaptation in order to allow for reliable real-time moderation systems. Through the episode-level and empirical trials, the approach attains effective outcomes via user-disjoint, temporally-separated validation. Thus, the framework achieves 96.5% recall, 98.5% total accuracy, 97.0% precision, 96.7% F1-score, and less than 2.0% false-positive rate. All of these attainments exhibit the proposed performance targets.