Abstract
An important problem in introducing personalization to query-based search retrieval and ranking is how to balance the relevance of results to the query with the affinity of the results to the user as oftentimes there is a trade-off between the two. This disclosure describes techniques to train a machine learning model to personalize ranking of search results in a controllable manner. With user permission, training data is constructed by analyzing query logs to identify manual query refinements followed by "long clicks," forming triplets of {original query, manually refined query, click}. These reflect personalized search results that the user found of high quality. Clicks from the original query are also obtained, reflecting other results for the query that were viewed. A machine learning model is trained using Softmax listwise loss to rank the true click from the manually refined query above clicks from the original query. The degree of personalization versus relevance can be controlled by tuning the temperature of the Softmax loss. The described techniques can be used in various search contexts.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Lin, Eric; Berlowitz, Devora; and Guo, Olive, "Machine Learning Driven Ranking to Control Degree of Personalization in Search Results", Technical Disclosure Commons, (August 06, 2025)
https://www.tdcommons.org/dpubs_series/8435