When the Group Agrees: Explaining Collective Recommendations From Core to Consensus

Finding an appropriate explanation for generated recommendations has become vital for increasing trustworthiness in current recommender systems (RSs). While the generation of explanations for recommendations aimed at individuals has recently received further study, explanations for group recommender systems (GRSs) scenarios remain inadequate for trustworthy GRSs. Furthermore, the majority of research on explaining group recommendation has predominantly focused on the user study evaluation stage by measuring the impact of explainable models in group members. Therefore, there remains a clear gap concerning frameworks and methodologies for systematising GRSs explanations. To address this, this paper adopts the feasibility of the post hoc explanation paradigm in GRSs, by discussing and analysing the use of four novel approaches supported by group members' preferences. Such approaches are: (1) CoreExplain, focused on generating explanations from a partial group cohesion; (2) CollectiveExplain, which considers the full group coverage for generating explanations; (3) HarmonyExplain, based on the shared individual evidence across explanations; and (4) ConsensusExplain, which unifies individual feature‐based explanations into a coherent group explanation. An offline experimental study, based on fidelity, coverage and diversity metrics, extensively evaluates the proposed approaches and includes the comparison with the state‐of‐the‐art. Furthermore, a user study is also developed for measuring satisfaction, perceived consensus, and the perceived fairness of the proposed approaches. Overall, the best performance was associated with CollectiveExplain and ConsensusExplain, and was closely linked to the internal characteristics of each associated dataset. Strengths and weaknesses of each approach were subsequently discussed.