Using Research Literature to Generate Datasets of Implicit Feedback for Recommending Scientific Items

Abstract

In an age of information overload, we are faced with seemingly endless options from which a small number of choices must be made. For applications such as search engines and online stores, Recommender Systems have long become the key tool for assisting users in their choices. Interestingly, the use of Recommender Systems for recommending scientific items remains a rarity. One difficulty is that the development of such systems depends on the availability of adequate datasets of users' feedback. While there are several datasets available with the ratings of the users for books, music, or films, there is a lack of similar datasets for scientific fields, such as Astronomy and Life and Health Sciences. To address this issue, we propose a methodology that explores scientific literature for generating utility matrices of implicit feedback. The proposed methodology consists in identifying a list of items, finding research articles related to them, extracting the authors from each article, and finally creating a dataset where users are unique authors from the collected articles, and the rating values are the number of articles a unique author wrote about an item. Considering that literature is available for every scientific field, the methodology is in principle applicable to Recommender Systems in any scientific field. The methodology, which we call LIBRETTI (LIterature Based RecommEndaTion of scienTific Items), was assessed in two distinct study cases, Astronomy and Chemistry. Several evaluation metrics for the datasets generated with LIBRETTI were compared to those derived from other available datasets using the same set of recommender algorithms. The results were found to be similar, which provides a solid indication that LIBRETTI is a promising approach for generating datasets of implicit feedback for recommending scientific items.

Keywords

• Recommender systems
• collaborative filtering
• scientific literature
• dataset
• astronomy
• chemical compounds