Multiset-driven attribute reduction for partially labeled hybrid data using fuzzy evidence theory and Gaussian kernel

As a generalization of traditional sets, a multiset can be used to represent object data, allowing certain data to appear multiple times. Capturing their frequencies can more accurately describe and analyze these data. This study proposes a multiset-driven approach for attribute reduction in partially labeled hybrid data using fuzzy evidence theory and Gaussian kernel. In a multiset decision information system label frequencies are used to represent known decisions and unlabeled objects are modeled as probability distributions over classes. Hellinger distance is employed to compute distances between objects and construct the fuzzy T cos -equivalence relation via Gaussian kernel. Fuzzy evidence theory defines fuzzy belief and plausibility maps based on fuzzy lower and upper approximations and designs two heuristic reduction algorithms. Experimental results on 15 datasets demonstrate that these algorithms outperform five existing methods in classification tasks and accuracy, highlighting the effectiveness of the multiset-driven framework in handling partially labeled hybrid data.