Training set reduction using Geometric Median

Learning large-scale dataset takes excessive processing time. Hence, smaller size of training set is beneficial to reduce the learning load. In this paper, a set of Geometric Medians are used as representative instances of the whole training set. Our proposed method can reduce the size of training sets to 0.015% - 10.81% of the original training set while the performance difference (F-Measure) is not over 6% from baseline models. In addition, this method has provided 1.91x to 7.01x speedup in learning time over the baseline models.