Simple Personalized Search Based on Long-Term Behavioral Signals

Extensive research has shown that content-based Web result ranking can be significantly improved by considering personal behavioral signals (such as past queries) and global behavioral signals (such as global click frequencies). In this work we present a new approach to incorporating click behavior into document ranking, using ideas of click models as well as learning to rank. We show that by training a click model with pairwise loss, as is done in ranking problems, our approach achieves personalized reranking performance comparable to the state-of-the-art while eliminating much of the complexity required by previous models. This contrasts with other approaches that rely on complex feature engineering.