Personalized Flight Itinerary Ranking at Fliggy

Abstract

Flight itinerary ranking is critical for Online Travel Agencies (OTAs) since more and more customers book flights online. Currently, most OTAs still adopt rule-based strategies. However, rule-based methods are not able to model context-aware information and user preferences. To this end, a novel Personalized Flight itinerary Ranking Network (PFRN) is proposed in this paper. In PFRN, a Listwise Feature Encoding (LFE) structure is proposed to capture global context-aware information and mutual influences among inputs. Besides, we utilize behaviors of both individual user and group users sharing the same intention to express user preferences. Then a User Attention Mechanism is proposed to rank flight itineraries based on the user preferences effectively and efficiently. Offline experiments on real-world datasets from Amadeus and Fliggy show the superior performance of the proposed PFRN. Moreover, PFRN has been successfully deployed on online system for searching itineraries at Fliggy and achieved significant improvements.