Semi supervised K–SVCR for multi-class classification

Abstract

In recent developments, the traditional binary class SVM has evolved into a multi-class classifier utilizing a ‘1-versus-1-versus-rest’ approach named K−SVCR. This innovative version efficiently categorizes multi-class data samples, generating ternary mode outputs $\{-1, 0, +1\}$, enabling simultaneous classification into three distinct classes. However, the availability of labelled data is scarce as it requires human effort for labelling it. But, the availability of unlabelled samples is easier as these are devoid of explicit class labels. This unlabelled data can be utilized by augmenting graph Laplacian regularisation term, evolving it into a semi-supervised learning approach for mode training. Thus, the model can capture samples’ distribution information available with abundant unlabeled data. Now, even a few labelled samples can elevate the effectiveness of the trained model on out-of-sample data. This article’s primary objective is to harness the potential of unlabeled data. To understand the intuition behind the proposed model two artificial datasets are considered. Further, the experiments were conducted on several real-world datasets from the UCI machine learning repository and the Flavia dataset for plant identification from leaf images, to evaluate the performance of the Semi-supervised graph Laplacian K−SVCR (Lap−KSVCR) model in terms of accuracy, precision, recall, and F-score of classification. It was observed that the model could achieve $63\%$ accuracy by training on a single labelled sample per class in case of non-linear three class artificial datasets. Additionally, on UCI real-world dataset and Flavia dataset an improvement of up to $17.5\%$, $17.3\%$, $30.1\%$, and $17.2\%$ was observed in accuracy, precision, recall, and f-score, respectively when compared to K−SVCR.