Enhancing the Multi-Objective Recommendation from three new perspectives: data characterization, risk-sensitiveness, and prioritization of the objectives

Abstract

Recommender Systems are tools whose main objective is to help users find relevant items among many options. However, different "relevance" concepts can be defined, making the recommendation task even more challenging if we want good recommendations on multiple quality concepts, e.g., accuracy, novelty, and diversity. In this scenario, the recommendation needs to use multi-objective optimization mechanisms. Although we find works focused on this type of recommendation, most of them are limited in some relevant aspects. In particular, three aspects provide scope for improving the multi-objective recommendation on new perspectives with the use of additional resources: (a) meta-features: implicit characteristics of input data can influence algorithms, e.g., quantity and distribution of items' ratings, therefore, explicit use of statistical measures capable of measuring some of those characteristics can be helpful in the multi-objective recommendation; (b) risk sensitivity: the optimization by global averages of multiple criteria can generate bad results in exchange for some excellent results that, although rare, can positively affect these averages, therefore, explicit use of risk sensitivity metrics can be helpful in the optimization process, reducing harmful recommendations without degrading global averages; (c) prioritization of objectives: users have different preferences regarding the quality criteria of recommendations, e.g., while some users do not give up favorite items, others may be more tolerant of discovering new items or a greater diversification of items, therefore, explicit use of users' preferences regarding the quality criteria can also be helpful to improve multi-objective recommendations further. Accordingly, in this work, we investigated the multi-objective recommendation from these three new perspectives and defined specific recommendation methods. Extensive experiments validated these methods, answered our research questions positively, and improved our knowledge concerning multi-objective recommendations on these three aspects, opening opportunities for relevant future work.