Learning to Recommend Related Entities With Serendipity for Web Search Users

Entity recommendation, providing entity suggestions to assist users in discovering interesting information, has become an indispensable feature of today’s Web search engine. However, the majority of existing entity recommendation methods are not designed to boost the performance in terms of serendipity, which also plays an important role in the appreciation of users for a recommendation system. To keep users engaged, it is important to take into account serendipity when building an entity recommendation system. In this article, we propose a learning to recommend framework that consists of two components: related entity finding and candidate entity ranking. To boost serendipity performance, three different sets of features that correlate with the three aspects of serendipity are employed in the proposed framework. Extensive experiments are conducted on large-scale, real-world datasets collected from a widely used commercial Web search engine. The experiments show that our method significantly outperforms several strong baseline methods. An analysis on the impact of features reveals that the set of interestingness features is the most powerful feature set, and the set of unexpectedness features can significantly contribute to recommendation effectiveness. In addition, online controlled experiments conducted on a commercial Web search engine demonstrate that our method can significantly improve user engagement against multiple baseline methods. This further confirms the effectiveness of the proposed framework.