Learning from Sets of Items in Recommender Systems

Reviewer: Brijendra Singh

Recommender systems are sets of computer algorithms or methods, implemented to provide suggestions or recommendations of relevant items to users. With the intensification of web services, from e-commerce to online advertising, recommender systems have become inevitable while using these online services. Most of the collaborative filtering methods utilize item-based past preferences provided by users. This paper explores an additional source of preferences "provided by users on sets of items," for example, ratings on a complete music album. Further investigation is done to describe "the user behavior related to rating sets [of items]" and "item-level rating predictions." Due to restrictions or privacy, users provide set-level ratings, but this mechanism does expose some user preference for many items. Apparently, this research evidently focuses on how a user's item-based preference conveys to their whole set-level preference, and how the existing item-based collaborative filtering model can benefit from such set-level ratings. Various models "for predicting the ratings that users will provide to the individual items" are discussed in detail, as well as how to "use these item-level ratings to derive set-level ratings." Model learning algorithms are very well defined. The paper includes a thorough statistical analysis, and the list of references is comprehensive. The authors use graphs, figures, and formulas to demonstrate their area of research. Performance testing and analysis of the proposed methods is performed on synthetically generated and real datasets from a popular online movie recommender system. This novel study is worthwhile to consider when enhancing existing recommender algorithms. This research is beneficial for intermediate and expert engineers, but beginners may have trouble understanding the complexity. It is necessary for all levels of engineers to be well versed in mathematical concepts.