Deducing window size and weights of Parzen classifier, using variogram function

Read full text Summarise with VIRAT
Authors: Pedram Masoudi
Cited by: 0
References: 25

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The Parzen window classifier assigns a class label to an unlabeled observation based on the similarity to known observations (majority voting). This method requires the optimization of two parameters: the Parzen window radius and kernel function. These two parameters are often optimized subjectively. In this article, a variogram function is introduced to achieve objective optimization. The variogram expresses the same spatial dependence as the covariance function but in terms of dissimilarity rather than similarity. The advantage of variogram function over covariance is in non-stationary condition where the covariance at high distances reveals correlations higher than that of distance zero, which is paradoxical. The variogram quantifies the variability as a function of distance in the feature space, allowing the definition of a geostatistical kernel function for each class-feature pair. The Euclidean distance is replaced by a variogram value, which is a measure of variability that reflects data heterogeneity. Moreover, the variogram-based Parzen window classifier does not require feature scaling or feature reduction, which are typically the necessary steps in classical algorithms. The proposed improvement in pattern recognition was tested on engineering-related or biological identification tasks (seven publicly available datasets), and it showed a high classification accuracy (from 2% to 80% improvement compared to the classical Parzen classifier). The accuracy of the Parzen-variogram was compared with two classifiers: the classical Parzen and k -nearest neighbor (KNN). The proposed method outperformed the classical Parzen method, and its accuracy was higher than that of the KNN method in two out of seven cases. In fact, the application of the Parzen-variogram classifier is limited to datasets with structured variogram models, which are used instead of Euclidean distance for class separation across features.

Fields of Study & Subjects

Data Mining and Machine Learning Applications Statistical Methods and Applications Machine Learning and Data Classification

Publication details

Journal / source
Machine Learning Engineering
Volume
2
Issue
1
Pages
015008–015008
Publication date
27 March 2026
Language
EN