Probabilistic Preference Learning with the Mallows Rank Model for Incomplete Data
Personalized recommendations are useful to assist users in their choices in web-based market places, entertainment engines, information providers. Learning individual preferences is an important step. Users express their preferences by rating, ranking, (possibly inconsistently) comparing, liking and clicking items. Such data contain information about the individual users ranking of the items. Click-through data can be seen as (consistent) pair comparisons. The Mallows rank model allows to analyse rank data, but its computational complexity has limited its use to a particular form, based on Kendalls distance. We developed new computationally tractable methods for Bayesian inference in Mallows models that work with any right-invariant distance. Our method performs inference on the latent consensus ranking of all items and on the individual latent rankings by Bayesian augmentation. Current popular recommendation algorithms are based on matrix factorisations, have high accuracy and achieve good clickthrough rates. However diversity of the recommended items is often poor and most algorithms do not produce interpretable uncertainty quantifications of the recommendations. With a simulation study and real life data examples, we demonstrate that compared to matrix factorisation approaches, our Bayesian Mallows method makes personalized recommendations mpared to matrix factorisation approaches, our Bayesian Mallows method makes personalized recommendations.
Arnoldo Frigessi is professor of statistics at the University of Oslo, leads the Oslo Center for Biostatistics and Epidemiology and is director of BigInsight. BigInsight is a centre of excellence for research-based innovation, a consortium of industry, business, public actors and academia, developing model based machine learning methodologies for big data. Originally from Italy, where he had positions in Rome and Venice, he moved to Norway in 2019 as a researcher at the Norwegian Computing Centre, before he became professor at the University of Oslo.
Frigessi has developed statistical methodology motivated by specific problems in science, technology and industry. He has designed stochastic models to study principles, dynamics and patterns of complex dependence. Inference is usually based on computationally intensive stochastic algorithms. Currently, he has research collaborations in genomics, personalised therapy in cancer, infectious disease models, eHealth research, personalised and viral marketing, s