Attribute Reduction for Incomplete and Unsupervised Data Based on Fuzzy Rough Sets

The high price and cost of obtaining data labels makes a lot of data in real world without labels. However, most of the existing attribute reduction approaches rely on decision attributes for research. In this situation, an attribute reduction approach is designed for unlabeled and incomplete information systems, which motivation is to explore the relationship between conditional attributes. Firstly, we introduce the calculation method of fuzzy similarity between samples, which is based on Gaussian kernel. Then the correlation between attributes is obtained, which can be used to evaluate the amount of information carried by the attribute subsets, that is, the significance of the attribute subsets. Next, we construct an attribute reduction algorithm called SUIFR based on fuzzy rough sets. Finally, using 6 data sets downloaded from UCI and KEEL to compare the clustering effect in k-means before and after data sets reduction, the experimental outcomes manifest the SUIFR has strong applicability in real world.