Preference Learning Under Sparse and Noisy Feedback in IIoT ‐Driven Industry 5.0 Systems: A Noise‐Corrected Graph‐Regularized Approach

Industrial Internet of Things (IIoT)‐driven Industry 5.0 systems increasingly rely on preference‐based decision modules to prioritize alarms, tasks, and actions under limited supervision. In practice, such feedback is both sparse and noisy due to constrained acquisition budgets and unreliable human–machine interaction channels, leading to biased and unstable ranking estimation. This paper studies preference learning under simultaneous sparsity and label corruption and proposes a unified framework that integrates noise‐corrected risk estimation with graph Laplacian regularization under a Bradley–Terry–Luce logistic model. The proposed objective explicitly debiases symmetric label noise through an ‐dependent correction while leveraging comparison topology to stabilize learning via structure‐aware smoothing. The resulting formulation remains compatible with stochastic gradient optimization and admits convergence guarantees in the convex setting. Experiments on three public ranking benchmarks under controlled sparsity and noise demonstrate consistent improvements in NDCG@10, MAP@10, and Kendall's over competitive baselines. Results confirm that jointly modeling corruption and comparison topology yields more robust and stable rankings, providing a principled approach for preference‐based decision modules operating under IIoT‐style supervision constraints.