WDset: A Semantic Distance Measure for Imbalanced Datasets

摘要

The concept of similarity/dissimilarity can be used to solve classification problems in imbalanced relational datasets where each record is a set of values corresponding to different attributes. Here, each attribute in the dataset has semantics, and therefore, it has a different effect on determination of category of the records. The distance measures that are commonly used to calculate similarity of the records compromise the semantics of the attributes rather than preserving it. In this paper, we propose a semantic distance measure named WDset which is a set of weighted distance components from two records. Here, individual distance components are calculated and are assigned a weight based on their relevance. Individual distance components preserve the semantics and weight handles the imbalanced nature of the dataset. Experimental tryouts show that the proposed measure is able to identify similar records more appropriately as compared to the existing measures.