Interval-value replacement-based neighborhood rough set model in incomplete fuzzy information systems

With the development of big data, incomplete fuzzy information systems (IFISs) exist in many applications. The processing of incomplete information (missing values) is an essential issue in the study of IFIS. Existing studies either increase the uncertainty of missing values, e.g., the neighborhood tolerance relation, or discard the uncertainty of missing values completely, e.g., the imputation approaches based on attribute relevancy. They may lead to unreasonable classification results. In this paper, we propose an interval-value replacement-based neighborhood rough set model (IVR-NRSM) from the perspective of preserving uncertainty to some extent rather than two extremes. According to two semantics of missing values, we first replace lost values in IFIS with interval values. Then the IFIS can be transformed into a replaced IFIS with only one semantic (i.e., the do not care). In the replaced IFIS, we define a distance function for numerical data and interval-value data. Furthermore, we construct the improved neighborhood tolerance relation and the corresponding neighborhood tolerance classes in the replaced IFIS. Finally, we design two experiments on 4 UCI data sets by introducing three performance metrics. Experimental results illustrate that the proposed IVR-NRSM has higher classification performance than the two representative models.