Title: Classification using distance nearest neighbours with adjusted pseudolikelihood
Authors: Lida Fallah - University College Dublin (Ireland) [presenting]
Nial Friel - University College Dublin (Ireland)
Abstract: The distance nearest neighbour (DNN) model, offers a probabilistic classification algorithm, modelling the joint distribution of the training and test data as a Markov random field. The essence of the DNN model is to account for the distance, in some sense, between feature vectors so that two feature vectors which are closer to each other are more likely to share the same class label. However, in its original formulation, it is computationally expensive due to an intractability of the likelihood. The pseudolikelihood offers a tractable alternative, however it is well understood that this can result in biased parameter estimation. To address this we consider a transformed pseudolikelihood approximation so that its mode and curvature (at the mode) coincide with that of the intractable likelihood. An additional advantage of our approach is it offers the possibility to carry out feature selection using Bayesian model selection.