Robust Partially-Observed VIoT Data Sensing via Half Quadratic Loss with Flexible Weighted Groupwise Relaxed Label Margins
Bo-Wei Chen · IEEE Internet of Things Journal · 2025
One distinctive feature of the Visual Internet of Things (VIoT) is its ability to prescreen and tag sensing data for efficient distribution to dedicated edge nodes. However, when partially observed data are captured, existing fitting models could become oversensitive. Besides, confining corrupted data to rigid and fixed categorical variable space may lead to biased outcomes, as equally spaced categorical variables could inadequately reflect the underlying distribution of corrupted data. Although state-of-the-art methods devised relaxed fixed categorical variables for enhancing performance, their relaxation strategies failed to consider same-class information. This led to inconsistent adjustments among samples within the same class. Consequently, the adjustments of the margin for the same class could potentially conflict. Besides, those methods were all based on convex$\ell _{2}$-norm loss functions, which limited the robustness against outliers. In contrast, existing robust approaches primarily focused on data modeling through nonconvex loss formulations but did not incorporate relaxed label margins, subsequently leaving the issue of flexible label margins unaddressed. Furthermore, to conquer the above problems, this study proposes robust data sensing based on nonconvex half quadratic loss for tackling oversensitivity stemming from partially observed data. To lift the constraint caused by rigid label variables and to increase regularization capabilities for label marginal space, flexible groupwise relaxation based on half quadratic loss is developed. It provides scaling and translation controls over marginal space. Additionally, class information is incorporated to reshape the marginal space during marginal formation by introducing groupwise dragging variables. The optimization procedure for groupwise dragging variables is specifically designed to account for the nonconvex nature of the half quadratic loss. Moreover, this study also derives flexible groupwise relaxation on Hamming space to accommodate the representation problem of partially observed input data. Experiments on VIoT data showed that the proposed method enhanced F1 scores by at least$\mathbf {6.67}\boldsymbol {\%}$,$\mathbf {8.45}\boldsymbol {\%}$, and$\mathbf {5.51}\boldsymbol {\%}$with respect to various forms of data corruption, including impulse noise, continuous occlusion, and missing values. This has verified the effectiveness of the proposed method.