Drift-Aware Cyberbullying Detection Using Bidirectional Mamba-Kan With Conformal Moderation

Abstract

Here, we propose a selective state-space sequence modelling and nonlinear spline based classification approach to detect cyberbullying. The suggested architecture is a hybrid of a bidirectional Mamba selective state-space encoder and a KAN classifier. The informal spelling and the abusive variants of words are preserved by byte-level BPE tokenisation, while forward and backward streams of the Mamba allow the representation of long conversational dependencies in a linear sequence of time, and gated fusion produces the representation input to the KAN decision layer (spline). The two additional extensions address the issue of reliability in deployment contexts: split conformal prediction provides set-valued moderation judgements and defers moderation in contexts of uncertainty and triggers threshold recalibration when online language drift is detected by ADWIN. A typical evaluation on the Multi-Label Twitter Dataset about Online Abuse achieves 97.6% accuracy, 97.8% precision, 97.4% recall, 97.6% F1-score, 0.952 MCC and 0.991 ROC-AUC, outperforming a 96.0% reference baseline. The resulting system is able to achieve high classification accuracy, uncertainty-aware moderation of evolving online language, and drift-aware maintenance.