Performance Evaluation of Multi-instance Multi-label Classification using Kernel based K-Nearest Neighbour Algorithm

A new learning methodology Multi-instance Multi-label learning (MIML) developed to address the problems of identifying multiple instances belonging to many class labels. This framework is used to understand how multiple instances or complex objects with different semantic meanings are related to many labels. In existing approaches algorithms like Support Vector Machines, AdaBoost, K-Nearest Neighbor, are adapted for the MIML framework. In this paper, a varied approach of adapting Polynomial Kernel and Linear Kernel before applying K-Nearest Neighbour algorithm for MIML framework is proposed. The Kernel provides a better linear separation of the data in the projected space and removes noise. Hausdorff metric is used in K-Nearest Neighbour algorithm to compute the distance between the training sample and the test sample. On performance evaluation it can be seen that the integration of the Polynomial Kernel and Linear Kernel method with the K-Nearest Neighbour algorithm for MIML framework (MIML-KKNN) has better performance in terms of ranking loss, and coverage errors when compared to KNN without Kernel function (MIML-KNN).