MKAN-Refine: Fine-Grained Crisis Information Mining via Reliability-Aware Nonlinear Refinement and Kolmogorov–Arnold Networks
Zheng Wang, Shanshan Li, Qingjie Liu, Zhian Pan, Xiaoling Sun · IEEE Access · 2026
Accurately mining fine-grained humanitarian information from social media streams is critical for crisis management but remains challenging due to the high semantic overlap between closely related categories and the heterogeneous reliability of multimodal data. Existing methods predominantly rely on Multi-Layer Perceptrons (MLPs) for feature fusion, creating a “Linear Bottleneck” that limits the disentanglement of complex topological structures in noisy environments. To address these limitations, this paper proposes MKAN-Refine, a novel architecture that integrates Kolmogorov–Arnold Networks (KANs) for reliability-aware nonlinear feature refinement. Unlike conventional linear projections, MKAN-Refine leverages learnable spline-based activations to model complex nonlinear interactions between visual and textual modalities. The framework incorporates a frozen CLIP backbone for stable embedding extraction, a dual-stream nonlinear attention module to enhance semantic discrimination, and a reliability-aware gating mechanism that adaptively balances modality contributions based on informational density. Extensive experiments on the CrisisMMD v2.0 benchmark demonstrate that MKAN-Refine achieves a Weighted F1-score of 92.07%, consistently outperforming strong CLIP-based baselines. Qualitative analysis further confirms that the proposed nonlinear refinement successfully resolves semantic ambiguities and establishes precise decision boundaries, offering a robust solution for automated crisis computing.