On the Pros and Cons of Explanation-Based Ranking

摘要

In our increasingly algorithmic world, it is becoming more important, even compulsory, to support automated decisions with authentic and meaningful explanations. We extend recent work on the use of explanations by recommender systems. We review how compelling explanations can be created from the opinions mined from user-generated reviews by identifying the pros and cons of items and how these explanations can be used for recommendation ranking. The main contribution of this work is to look at the relative importance of pros and cons during the ranking process. In particular, we find that the relative importance of pros and cons changes from domain to domain. In some domains pros dominate, in other domains, cons play a more important role. And in yet other domains there is a more equitable relationship between pros and cons. We demonstrate our findings on 3 large-scale, real-world datasets and describe how to take advantage of these relative differences between pros and cons for improved recommendation performance.