Institutional Scholarship

Recommender System for Scientific Explorations

Show simple item record

dc.contributor.advisor Friedler, Sorelle Yu, Vincent 2021-07-12T11:57:42Z 2021-07-12T11:57:42Z 2021
dc.description.abstract Conventional recommender systems tend to mainly focus on recommendation accuracy and are generally optimized to achieve the maximum predicted recommendation relevance. However, as many researchers now noticed, accuracy merely captures a fraction of the recommender system's performance, and ignoring other aspects of the recommendations, such as diversity and novelty, can hinder the overall quality and usability of the recommender systems. This issue, often named the "accuracy overloading problem", becomes more complicated in the domain of scientific explorations, where recommender systems can theoretically provide a sensible leap into the realm of the unknown. However, different projects and researchers may have different opinions on what types of leaps are desirable, and how big of a leap is acceptable. Hence, recommender systems used for scientific explorations must explicitly incorporate the beyond-accuracy objectives and carefully fine-tune the balance between recommendation accuracy and these objectives with the support of reasonable evaluation metrics. To resolve the "accuracy overloading" problem and design a suitable recommender system for scientific explorations, this paper introduces a top-K dual-objective recommender system framework that optimizes both recommendation relevance and beyond-accuracy qualities simultaneously. And to demonstrate the performance of the proposed framework and compare its difference from a traditional accuracy-focused recommender system, the dual-objective recommender system is then implemented to recommend chemical reactions with high serendipity values for the Dark Reactions Project. After several offline experiments for validation and live experiments for testing, the proposed recommender system is proven to be a significant improvement over the traditional accuracy-focused recommender system, raising recommendation serendipity from nearly 0 to 0.33 and recommendation accuracy from 30% to 70% at the same time.
dc.description.sponsorship Haverford College. Department of Computer Science
dc.language.iso eng
dc.title Recommender System for Scientific Explorations
dc.type Thesis
dc.rights.access Bi-College users only until 2022-01-01, afterwards Open Access.

Files in this item

This item appears in the following Collection(s)

Show simple item record Except where otherwise noted, this item's license is described as



My Account