Abstract
•Proposes the Kernel-based dynamic online classification algorithm for detecting fake reviews in dynamic environments.•Enhances model robustness against noise and outliers with a slope-adjusted ramp loss function.•Theoretical analysis establishes its generalization guarantees and noise resistance.•Validates the algorithm’s performance on real-world datasets from TripAdvisor, Yelp, and Amazon.
Online consumer reviews are pivotal in e-commerce, yet their integrity is increasingly threatened by sophisticated fake reviews that undermine market trust and consumer confidence. The dynamic, noisy, and unstructured nature of this user generated content poses significant challenges for conventional detection methods. To address this, we propose Dynamic Classification for Online Content (DCOC), a kernel-based online framework for real-time data streams that incorporates a slope-adjusted ramp loss to enhance robustness against noise and outliers. Through extensive experiments on real-world datasets from TripAdvisor, Yelp, and Amazon, DCOC delivers superior or highly competitive performance across datasets, with statistically significant gains over most baselines. Further noise tolerance experiments confirm its stability, maintaining strong performance even with up to 30 % data noise. These results validate DCOC as a robust, adaptive, and efficient solution for detecting fake reviews in complex, dynamic online environments.