A Comparative Analysis of Convergence Rate for Imbalanced Datasets of Active Learning Models

Currently, active learning is widely used in several applications to build an initial classification system for the trivial amount of data sets. The main problem of the modern active learning systems is their assumption that the training sets are perfect, and they don’t take into consideration the data issues derived from the real-world scenarios. These research challenges of existing models could cause several concerns in the real-time application including redundancy, and incoherence or the big size of data among others. In this research, we compared the six active learning methods based on nine standard datasets derived from UCI database in terms of their convergence rate. From the experimental results, it is observed that these methods don’t achieve high performance of learning due to the convergence rate or information loss. The comparative analysis based on nine test reveals that the decision-tree based active learning method produces seven times optimal convergence rate for imbalanced data with notable sample attribute difference.