2023-07-10 ニューヨーク大学 (NYU)
◆しかし、コンピュータサイエンティストのチームが開発したPyrorankアルゴリズムは、生態系の相互作用を模倣して、推薦の幅を広げることができます。このアルゴリズムは、ユーザーのプロフィールの影響を減らし、より多様な結果を提供します。既存の推薦システムと比較して、Pyrorankはより多様な推薦を生成し、フィルターバブルを打破する価値があることが示されています。さらに、特定の使用ケースに応じて調整が可能であり、推薦の多様性を高めながら予測の正確性を保つことができます。このアルゴリズムは、既存の推薦システムの制限や偏りを克服し、より効果的な推薦とプラットフォームの長期的な健全性を促進する重要な一歩となります。
<関連情報>
- https://www.nyu.edu/about/news-publications/news/2023/july/researchers-devise-algorithm-to-break-through–search-bubbles-.html
- https://link.springer.com/chapter/10.1007/978-3-031-36625-3_12
Pyrorank: レコメンダーシステムの多様性を促進する、自然から着想を得た新しいアルゴリズム Pyrorank: A Novel Nature-Inspired Algorithm to Promote Diversity in Recommender Systems
Doruk Kilitcioglu,Nicholas Greenquist & Anasse Bari
Advances in Swarm Intelligence published:08 July 2023
DOIh:ttps://doi.org/10.1007/978-3-031-36625-3_12
Abstract
Recommender systems are essential to many of the largest internet companies’ core products. Today’s online users expect sites offering a vast assortment of products to provide personalized recommendations. Although traditional recommender systems optimize for prediction accuracy, such as RMSE, they often fail to address other important aspects of recommendation quality. In this paper, we explore the crucial issue of diversity in the recommendations generated by recommender systems. We explain why diversity is essential in recommender systems and review related work on diversifying recommendations. We quantify and classify various diversity metrics into logical categories. Then, we introduce Pyrorank, a novel bio-inspired re-ranking algorithm designed to improve recommendation diversity. Pyrorank is inspired by the positive effects of pyrodiversity in nature and is optimized to increase user-based diversity and mitigate the systemic bias that traditional recommender system models learn from the data. Our experimental results on multiple large datasets indicate that Pyrorank can achieve better user-based diversity metrics than state-of-the-art re-ranking methods, with little decrease in prediction accuracy.