Multi-Instance Learning (MIL) By Finding an Optimal Set of Classification Exemplars (OSCE) Using Linear Programming

This paper describes how to classify a data set by using an optimum set of exemplars to determine the label of an instance among a set of data for solving a classification run time problem in a large data set. In this paper, these exemplars purposely have been used to classify positive and negative bags in a synthetic data set. There are several methods to implement multi-instance learning (MIL) such as SVM, CNN, and Diverse density. An optimum set of classifier exemplar (OSCE) is used to recognize positive bag (contains tumour patches). The goal of this paper is to find a way to speed up the classifier run time by choosing a set of exemplars. A linear programming approach is been used to optimize a hinge loss cost function, in which estimated label and actual label is used to train the classification. The estimated label is calculated by measuring the Euclidean distance of a query point to all of its k nearest neighbours and an actual label value. To select some exemplars with none zero weights, two solutions are suggested to have a better result. One of them is choosing k closer neighbours. The other one is using LP and thresholding to select some maximum of achieved unknown variable which are more significant in finding a set of exemplars. Also, there is a trade-off between classifier run time and accuracy. In a large data set, the OSCE classifier has better performance than ANN and K-NN cluster. Also, OSCE is faster than the NN classifier. After describing the OSCE method, this has been used to recognize a data set that contains cancer in synthetic data points. Eventually, the defined OSCE has been applied to MIL for cancer detection.