Generalized Mixed Effect Models for Personalizing Job Search

Job Search is a core product at LinkedIn which makes it essential to generate highly relevant search results when a user searches for jobs on Linkedin. Historically job results were ranked using linear models consisting of a combination of user, job and query features. This paper talks about a new generalized mixed effect models introduced in the context of ranking candidate job results for a job search performed on LinkedIn. We build a per-query model which is populated with coefficients corresponding to job-features in addition to the existing global model features. We describe the details of the new method along with the challenges faced in launching such a model into production and making it efficient at a very large scale. Our experiments show improvement over previous baseline ranking models, in terms of online metrics (both AUC and NDCG@K metrics) as well as online metrics in production (Job Applies) which are of interest to us. The resulting method is more powerful and has also been adopted in other applications at LinkedIn successfully.