Distributional learning in multi-objective optimization of recommender systems

Abstract Metrics such as diversity and novelty have become important, beside accuracy, in the design of Recommender Systems (RSs), response increasing users' heterogeneity. Therefore, RSs is now increasingly modelled a multi-objective optimization problem (MOP) for whose solution Multi-objective evolutionary algorithms (MOEAs) been considered. In this paper we focus on k-top recommendation which encoded matrix rows correspond to customers column items. The value novelty, coverage each candidate list, evaluated sample can be represented 3-d histogram encodes knowledge obtained from function evaluations. This enables map space into space, elements are histograms, structured by Wasserstein (WST) distance between histograms. similarity 2 users probabilistic given their construction WST graph nodes weights edges users. clustering takes then place WST-graph. phase difference two top-k lists 3-dimensional derive new selection operators provide better diversification (exploration). algorithm optimization/Wasserstein (MOEA/WST), compared with benchmark NSGA-II, yields hypervolume coverage, particular at low generation counts.