New Framework Optimizes Scale Reduction and Corrects Inconsistent Decision Labels

Researchers have developed a framework for label correction and dependency-degree optimal scale reduction to address knowledge acquisition challenges in inconsistent generalized multi-scale hybrid decision systems. Writing in the International Journal of Machine Learning and Cybernetics, authors Shi-Lin Zhou, Wei-Zhi Wu, Anhui Tan, and Harry F. Lee examined how erroneous or unreliable decision labels affect analytical processes in generalized multi-scale hybrid decision environments.

To address label inconsistencies, the investigators first constructed decision-distribution-based object distances for set-valued and nominal condition attributes, establishing corresponding similarity relations. Using the similarity classes induced by the complete condition attribute set under the strongest scale combination, the team proposed a label correction method based on maximum distribution decision classes.

With the corrected labels held fixed, the authors constructed similarity relations and rough approximations across various scale combinations. By relying on the dependency degree under the strongest scale combination, they introduced the concepts of dependency-degree optimal scale combinations and dependency-degree optimal scale reducts, which maintain the reference dependency degree while satisfying a minimality condition.

The researchers also formulated overlap measures for set-valued and nominal condition attributes, introducing a multi-scale average overlap degree to quantify attribute discriminability across multiple scales and dictate search order. An algorithm guided by the multi-scale average overlap degree was developed to search for dependency-degree optimal scale reducts. Subsequent comparative experiments on eight UCI datasets evaluated under CART and RBF-SVM classifiers demonstrated the effectiveness and validation of the approach.

According to the published study, knowledge acquisition from generalized multi-scale hybrid decision systems is frequently hindered when decision labels are erroneous or unreliable. To resolve this issue in inconsistent generalized multi-scale hybrid decision systems, the research establishes decision-distribution-based object distances specifically for set-valued and nominal condition attributes, leading directly to the formulation of corresponding similarity relations. The label correction procedure operates on the similarity classes induced by the full condition attribute set under the strongest scale combination, applying maximum distribution decision classes to revise the problematic labels.

Following the label correction step, the corrected decision labels are kept fixed while the investigators construct similarity relations and rough approximations under different scale combinations. The framework utilizes the dependency degree under the strongest scale combination as a reference benchmark. From this foundation, the authors introduce the formal concepts of dependency-degree optimal scale combinations and dependency-degree optimal scale reducts. These combinations and reducts are defined to preserve the reference dependency degree while satisfying an explicit minimality condition across the evaluated scales.

To establish the search order and assess attribute discriminability across multiple scales, the researchers construct overlap measures for both set-valued and nominal condition attributes. These measures form the basis of the proposed multi-scale average overlap degree. The team then incorporates this metric into an algorithmic search strategy, developing a dedicated multi-scale average overlap degree-guided algorithm to locate the dependency-degree optimal scale reduct.

The practical utility of the complete methodology was evaluated through comparative experiments across eight distinct UCI datasets. The testing procedures utilized CART and RBF-SVM classifiers to assess performance. The resulting experimental outcomes demonstrated both the effectiveness and the validation of the proposed label correction and dependency-degree optimal scale reduction approach in handling inconsistent generalized multi-scale hybrid decision systems.

Subject of Research: Label correction and scale reduction in inconsistent generalized multi-scale hybrid decision systems

Article Title: Label correction and dependency-degree optimal scale reduction for inconsistent generalized multi-scale hybrid decision systems

Article References: Zhou, S.-L., Wu, W.-Z., Tan, A., & Lee, H. F. (2026). Label correction and dependency-degree optimal scale reduction for inconsistent generalized multi-scale hybrid decision systems. International Journal of Machine Learning and Cybernetics, 17(12), Article 531. https://doi.org/10.1007/s13042-026-03358-0

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03358-0

Keywords: machine learning, rough set theory, decision systems, scale reduction, label correction, hybrid decision systems, attribute reduction, data classification, Label, correction, dependency-degree, optimal