Methods for Predicting the Rise of the New Labels from a High-Dimensional Data Stream

Abstract

In data engineering, multi-label learning (MLL) has emerged to classify the instances through a specific characteristic that associates with the set of class labels (CLs). Mostly, the learning design was adaptive and newer views may exist in a data stream (DS); so, MLL has to classify the features with newer CLs. To tackle this issue, the MLL with emerging new labels (MuENL) and handling high-dimensional DSs (MuENLHD) technique has been adopted which considers the CLs in the test data were similar to that in the learning data. But, it was not able to deal with the adaptive situation where multiple newer CLs exist since it can manage only a single newer CL in one iteration. Hence this article proposes an MLL with emerging multiple new labels (MuEMNL) and MuEMNLHD versions to combat the issues in a complex scenario wherein several newer CLs are found. The main idea of this technique is to split the newer CL group into many newer CLs independently for changing the complex scenario. In this technique, four different steps are executed: i) creates a linear classification model to adjust the pairwise CL sorting error and the categorization error on the given CLs, ii) develops a novel outlier identifier depending on the primary and test DS, iii) discovers the cluster for the MuEMNL and MuEMNLHD depending on the OPTICS clustering and iv) employs a classifier updating scheme to integrate newer CLs for designing a robust classifier. Finally, the experimental outcomes exhibit that these techniques on low-dimensional databases attain an overall mean precision of 67.08% and an overall F1-score of 64.02% compared to the classical MLL techniques. Similarly, these techniques on high-dimensional databases achieve mean precision of 64.6% and a mean F1-score of 63.1% compared to the classical MLL techniques.