Partially Observed Visual IoT Data Reconstruction Based on Robust Half-Quadratic Collaborative Filtering

Recent advances and applications in the Visual Internet of Things (VIoT) have witnessed significant progress in various areas, such as smart sensing and environmental monitoring, which rely on sensors to capture data to build analytic models. However, when VIoT devices are deployed in harsh environments, such as the deep ocean and underground, the physical integrity and functionality of these devices may deteriorate. As a result, there can be disruption in the data acquisition process, leading to incomplete data generated in the sink node owing to interference, occlusion, connectivity, or unknown abnormal events. Poor data quality affects the validity and generalizability of the analysis. Although numerous data recovery algorithms have been developed,
$\ell _{2}$
,
$\ell _{1}$
, and
$\ell _{2,1}$
norms were unable to effectively handle oversensitivity during data modeling. Despite the fact that recent half quadratic (HQ) functions have shown improvement in robustness, the models still degraded and produced fitting biases owing to nonflexible loss functions and rigid reweighting when sufficient large fitting errors occurred. In view of such, this study proposes collaborative filtering based on scaled lifted augmented HQ functions to accommodate the above-mentioned problems. First, additional auxiliary variables are introduced to flexibly absorb varying levels of data corruption through derived lifting schemes. Moreover, this study devises the scaled reweighting forms for the proposed lifted augmented HQ functions to further increase resilience against oversensitivity. Experiments on open data sets showed that the proposed method yielded better recoverability than the baselines under different data corruption conditions. Such findings verified the effectiveness of the proposed model.