2026/04/21 by Alberto Durán-López, Daniel Bolaños-Martinez, María Bermúdez-Edo · 1 voice
Computer Science · #Advanced Database Systems and Queries #Advanced Graph Neural Networks #Data Visualization and Analytics
paper · doi:10.1109/jiot.2026.3685702
openalex publication_date 2026/04/21 · openalex created_date 2026/04/22 · openalex updated_date 2026/08/01
Multiview learning in Internet of Things environments offers the potential to integrate complementary views to capture complex information dynamics. However, standard fusion mechanisms often fail when specific views exhibit uncertainty, as these mechanisms typically assign fixed importance to all views, disregarding that the information contained in each view may vary dynamically. To address this limitation, we propose Gated Cross-View Attention (GCVA), a view-agnostic fusion mechanism that estimates view-specific uncertainty and dynamically downweights views with high uncertainty before they corrupt the shared feature space. Then, GCVA integrates the uncertainty estimation with cross-view gated-attention mechanisms to fuse the views. We validate this approach using a real-world License Plate Recognition (LPR) dataset from a smart tourist village, constructing three aligned views: temporal, tabular, and graph, to predict vehicle revisits. We benchmark our proposal against established fusion mechanisms, including Gated Multimodal Units (GMU), attention and hybrid approaches. Results demonstrate that GCVA effectively modulates the influence of high-uncertainty views, compared to baseline mechanisms. Our proposal achieves a Weighted F1-score of 0.730, outperforming state-of-the-art baselines by a margin of 2.8% to 14.2%, while reducing training time by 4.7% to 40.2%. These findings confirm that explicit uncertainty modeling is important for robust forecasting in realworld IoT scenarios, particularly when the predictive reliability of the views varies.