Enhanced SMART-TV: A Classifier with Vertical Data Structure and Dimensional Projections Pruning

Classification is the process of assigning class membership to the newly encountered and unlabeled objects based on some notion of closeness between the objects in the training set and the new objects. In this work, we introduce the enhancement of SMART-TV (SMall Absolute diffeRence of ToTal Variation) classifier by introducing dimensional projections process to prune the neighbors in the candidate set that are far away from the unclassified object in terms of distance but are close in terms of total variation. SMART-TV is scalable nearest-neighbors classifier. It uses vertical data structure and approximates a set of potential candidate of neighbors by means of vertical total variation. The total variation of a set of objects about a given object is computed using an efficient and scalable Vertical Set Squared Distance algorithm. In our previous work of SMART-TV, the dimensional projections was not introduced, and thus, the candidate set that are far away from the unclassified object but close in terms of total variation also voted. This, in some cases, can reduce the accuracy of the algorithm. Our empirical results using both real and synthetic datasets show that the proposed dimensional projections effectively prune the superfluous neighbors that are far away from the unclassified object. In terms of speed and scalability, the enhanced SMART-TV is very fast and comparable to the other vertical nearest neighbor algorithms, and in terms of accuracy, the enhanced SMART-TV with dimensional projections pruning is very comparable to that of KNN classifier.