Aggregate Metric Networks for Nearest Neighbour Classifiers

A common idea prevailing in distance or similarity measures is the use of aggregate operators in localised and point-wise differences or similarity calculation between two patterns. We test the impact of aggregate operations such as min, max, average, sum, product, sum of products and median on distance measures and similarity measures for nearest neighbour classification. The point-wise differences or similarities extends the idea of distance measurements from the decision space to feature space for the extraction of inter-feature dependencies in high dimensional patterns such as images. Inter-feature spatial differences are extracted using the gradient functions across various directions and then applied on aggregate function, to result in a fused feature set. The initial study is conducted on Iris flower and verified using AR face database. The resulting method shows an accuracy of 92% on face recognition task using the standard AR database.