Learning to rank aspects and opinions for comparative explanations

Trung-Hoang Le, Hady W. Lauw

Abstract: Comparative recommendation explanations help to make sense of recommendations by comparing a recommended item along some aspects of interest with one or many items being considered. This work extends the notion of comparative explanations, by going beyond merely better/worse statements, to further incorporate aspect-level opinions for more informative comparisons. To enhance the quality of both the personalized recommendation and the explanation, we incorporate optimization objectives that preserve relative rankings of aspects and opinions, in addition to the classical rankings of overall preferences for items. We integrate the multiple ranking objectives and multi-tensor factorization together. Experiments on datasets of different domains validate the efficacy of our proposed framework in both recommendation and comparative explanation against comparable explainable recommendation baselines.