The binomial-neighbour instance-based learner on a multiclass performance measure scheme

This paper presents a novel instance-based learning methodology the Binomial-Neighbour (B-N) algorithm. Unlike to other k-Nearest Neighbour algorithms, B-N employs binomial search through vectors of statistical features and distance primitives. The binomial combinations derived from the search with best classification accuracy are distinct primitives which characterise a pattern. The statistical features employ a twofold role; initially to model the data set in a dimensionality reduction preprocessing, and finally to exploit these attributes to recognise patterns. The paper introduces as well a performance measure scheme for multiclass problems using type error statistics. We harness this scheme to evaluate the B-N model on a benchmark human action dataset of normal and aggressive activities. Classification results are being compared with the standard IBk and IB1 models achieving significantly exceptional recognition performance.