A0261
Title: A robust and adaptive $k$NN ensemble using sequential weighted neighborhood selection
Authors: Zardad Khan - United Arab Emirates University (United Arab Emirates) [presenting]
Shamsa Alnasri - United Arab Emirates University (United Arab Emirates)
Amjad Ali - United Arab Emirates University (United Arab Emirates)
Saeed Aldahmani - United Arab Emirates University (United Arab Emirates)
Abstract: Conventional $k$NN ensembles detect the $k$ nearest observations within a pre-determined spherical area identified by a distance metric. However, this procedure may suffer when the test instances follow a structured pattern adhered to its most related neighbors along a sequential trail, which the static neighborhood may fail to capture. To address this limitation, a $k$NN ensemble is proposed where neighborhood selection is carried out in a sequential fashion utilizing a weighted distance measure, where the weights are obtained using the differential capability of features into two classes. The neighborhood selection starts by finding the first nearest neighbor using the weighted distance, iteratively followed by identifying successive neighbors that are nearest to the formerly determined neighbor, continuing until $k$ neighbors are identified. For diversity and generalizability, each base learner in the ensemble uses a bootstrap sample with a randomly selected subset of features. Majority voting is used across all the base models to obtain the final predictions. The suggested method is assessed on 20 benchmark classification problems against conventional $k$NN-based ensembles and other state-of-the-art classifiers using various performance metrics. Experimental results reveal that the proposed ensemble outperforms the other methods in the majority of cases.
