Scalable recommendations using decomposition techniques based on Voronoi diagrams

Collaborative filtering based recommender systems typically suffer from scalability issues when new users and items join the system at a very rapid rate. We tackle this concerning issue by employing a decomposition based recommendation approach. We partition the users in the recommendation domain with respect to location using a Voronoi Diagram and execute the recommender algorithm individually in each partition (cell). This results in a much reduced recommendation time as we eliminate the need for running the algorithm using the entire user set. We further address the problem of improving the recommendation quality of the users residing in the peripheral region of a Voronoi cell. The primary objective of our approach is to bring down the recommendation time without compromising the accuracies of recommendations much, which is rightly addressed by our proposed method. The outcomes of the experiments performed demonstrate the scalability as well as efficacy of our method by reducing the runtime of the baseline CF algorithm by at least 65% for each of these four publicly available datasets of varying sizes - MovieLens-100K, MovieLens-1M, Book-Crossing and TripAdvisor datasets. The accuracies of recommendations in terms of MAE, RMSE, Precision, Recall and F1 metrics also hold good.