Two-stage Interest Calibration Network for Reranking Hotels

Abstract

As one key task of hotel search, interest modelling is to capture users’ interests over their historical clicks. To this end, it is important yet challenging to understand the searching intention behind each click (interest focus). Besides, the query input by a user represents the user’s interests to some extent. The differences between historical queries and the current query result in the gap between historical interests and the current interests (interest gap). Because historical clicks stem from historical interests, using them to model the current interests inevitably suffers from the interest gap challenge. To capture the interest focus and address the interest gap, we propose the two-stage interest calibration network (TCN), i.e., search-internal and search-external. In the search-internal calibration, we propose new insights of using the divergences among clicks and unclicks to model interest focus, and then develop a divergence-based calibration network. In the search-external calibration, inspired by the smoothing techniques for language models, we propose the interest smoothing principle to bridge interest gap: the interests learnt from historical clicks $+$ smoothing factor $\approx$ current interests. To implement this principle, we develop an interest smoothing network by reusing the query data. In the network, the interest domination unit is developed to learn user’s interests from historical clicks, and the interest smoothing unit is developed to construct the smoothing factor. Extensive offline experiments and online A/B testing are performed and show that TCN significantly outperforms the state-of-the-art baselines. Besides, our model has been deployed in a hotel e-commerce platform and brought $2.90\%$ CTCVR and $1.53\%$ CTR lifts.

D. Ma and J. Sun—Equal Contribution.