Presentation Information
[5Yin-A-23]Sales Uplift Prediction for Point-Based Promotion Campaigns in an Online Travel Agency
〇Jiheng Sun1, Hisashi Kashima1, Koh Takeuchi1, Kyohei Atarashi1, Ryo Matsui2, Shusaku Yoshizumi2, Shunji Umetani2 (1. Kyoto University, 2. Recruit Co., Ltd.)
Keywords:
Uplift Modeling,Zero-Inflated Model,Marketing Effect Prediction
In this study, we propose a user-level method to predict the effects of point-based promotion campaigns aimed at encouraging hotel reservations on an online travel agency (OTA) platform. While OTAs can implement efficient marketing strategies by adjusting whether and how many points are granted to each user, accurately predicting the effects of such campaigns in advance remains challenging.
To address this problem, we apply uplift modeling to predict the incremental increase in hotel reservation revenue resulting from point allocation. Reservation revenue has a zero-inflated distribution, as it becomes zero when no reservation is made. Taking this characteristic into account, we construct a zero-inflated uplift model that combines a model for predicting the probability of a reservation with a model for predicting the revenue amount conditional on a reservation occurring.
The proposed approach separately estimates the change in reservation probability and the change in revenue amount induced by point allocation, and integrates them to estimate the uplift in the expected reservation revenue. Experiments on real-world data demonstrate that, compared with conventional methods that directly predict uplift in reservation revenue, the proposed zero-inflated uplift model achieves improved uplift predictive performance.
To address this problem, we apply uplift modeling to predict the incremental increase in hotel reservation revenue resulting from point allocation. Reservation revenue has a zero-inflated distribution, as it becomes zero when no reservation is made. Taking this characteristic into account, we construct a zero-inflated uplift model that combines a model for predicting the probability of a reservation with a model for predicting the revenue amount conditional on a reservation occurring.
The proposed approach separately estimates the change in reservation probability and the change in revenue amount induced by point allocation, and integrates them to estimate the uplift in the expected reservation revenue. Experiments on real-world data demonstrate that, compared with conventional methods that directly predict uplift in reservation revenue, the proposed zero-inflated uplift model achieves improved uplift predictive performance.