Online academic repositories help millions of researchers discover relevant articles, a domain in which there are many potential signals of relevance, including text, citation links, and how recently an article was published. In this paper we present a case study of productionizing learning to rank for large scale recommendation, which utilises these diverse feature sets to increase user engagement. We first introduce item-to-item collaborative filtering (CF), then how these recommendations are rescored with a LtR model. We then describe offline and online evaluation, which are essential for productionizing any recommender. Nonetheless, the online results show that learning to rank significantly increased user engagement with the recommender. Finally we show through post-hoc analysis that the original CF solution tended to promote older articles with lower traffic. However, by learning from subjective user interactions with the recommender system, our relevance model reversed those trends.
|Original language||American English|
|State||Published - 2020|
|Event||European Conference on Information Retrieval - Lisbon, Portugal|
Duration: Apr 14 2020 → Apr 17 2020
|Conference||European Conference on Information Retrieval|
|Period||04/14/20 → 04/17/20|
Kernshaw, D., Pettit, B., Hristakeva, M., & Jack, K. (2020). Learning to Rank Research Articles: A case study of collaborative filtering and learning to rank in ScienceDirect. Paper presented at European Conference on Information Retrieval, Lisbon, Portugal.