Exploiting user preference similarity transitivity in nearest neighbour recommender algorithms

Abstract: This thesis explores the Nearest Neighbour recommender algorithm and a proposed way of recomputing the elements of a set, containing values from a similarity metric. In essence, this extends the neighbourhood used in the algorithm by making use of close neighbours’ neighbours. This proposal is motivated by the hypothesis that high user preference similarity is transitive. A programme was implemented in Matlab, and one of the MovieLens data sets was used. Three experiments were run. These indicated that further exploration of this property and the proposed methods could be of interest. The final experiment was run on the simplest form of a Nearest Neighbour recommender, with the added re-computational step, and suggests that there is a small, but statistically significant, improvement due to this new step (for one of the proposed methods) confirming the hypothesis. The improvement of the errors of estimated rating predictions were in the order of magnitude 0.1 %.