A class-driven approach to dimension embedding. (1st June 2022)

• Record Type: Journal Article
• Authors: Aydın, Fatih
• Is Part Of: Expert systems with applications. Volume 195(2022)
• Publication Date: 2022-06-01
• Subjects: Machine learning -- Dimension embedding -- Cumulative distribution function -- Brownian motion -- Concave-convex functions -- Distance metrics; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.116650
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: The proposed algorithm called CDE carries out an exact embedding for test points. CDE does not suffer from outliers due to utilizing class information well. A class-aware embedding provides more explanatory information about the classes. A new distance metric based on information gain is proposed. The embedded data delivers a higher accuracy than originals in many data sets. Abstract: The preservation of the inbuilt structures of data sets and the more decomposition of the classes are a significant interest in dimension embedding. In this respect, the dimensionality reduction methods use novel techniques to better ascertain the fundamental structure of the manifold on which the data lies. However, both conventional and state-of-art supervised dimensionality reduction methods cannot benefit from class information good enough. Therefore, their generalization performances on the test data are weak. A new non-linear supervised algorithm, which we call Class-driven Dimension Embedding (CDE), is proposed for utilizing class information. CDE performs three outstanding characteristics: (i) preserving the intrinsic relationship between the data points and classes; (ii) producing wide margins between classes; (iii) enhancing the generalization performance on the test data. The proposed method embeds a d-dimensional data set into the c-dimensional space (c designates the number of classes) through the corresponding values to classes of each point by exploiting a neighborhood graph and a feature weighting function. The experimental results on forty-eight data sets demonstrate that CDE is comparable to or better than twenty-four dimensionality reduction algorithms in terms of classification accuracy and visualization. The source code of CDE can be found in https://doi.org/10.24433/CO.0967299.v1 for computational reproducibility.