The Informative Vector Selection in Active Learning using Divisive Analysis
Zareen Sharf, Maryam Razzak · International Journal of Advanced Computer Science and Applications · 2017
Traditional supervised machine learning techniques require training on large volumes of data to acquire efficiency and accuracy. As opposed to traditional systems Active Learning systems minimizes the size of training data significantly because the selection of the data is done based on a strong mathematical model. This helps in achieving the same accuracy levels of the results as baseline techniques but with a considerably small training dataset. In this paper, the active learning approach has been implemented with a modification into the traditional system of active learning with version space algorithm. The version space concept is replaced with the divisive analysis (DIANA) algorithm and the core idea is to pre-cluster the instances before distributing them into training and testing data. The results obtained by our system have justified our reasoning that pre-clustering instead of the traditional version space algorithm can bring a good impact on the accuracy of the overall system’s classification. Two types of data have been tested, the binary class and multi-class. The proposed system worked well on the multi-class but in case of binary, the version space algorithm results were more accurate.