Personalized Demand Prediction and Ranking for Qunar's "Guess You Like" Feature

Many modern apps, such as Taobao, Amazon, and YouTube, rely on user demand prediction to recommend items; Qunar replaced a low‑click‑rate hot‑words card on its search homepage with a "Guess You Like" card to increase click‑through rate and reduce bounce.

To discover user‑demand associations, the system uses a rule engine, an A2B transition‑matrix collaborative filtering method, location‑based service (LBS) collaborative filtering, and dozens of manually crafted rules for special scenarios.

The ranking pipeline progressed from a subjective Bayes model (assuming independent expert opinions and discrete rank features) to a RankBoost model that automatically learns feature discretization, and finally to a LambdaMart model that combines LambdaRank and GBDT, leveraging tree structures to capture dependencies among rules.

Training samples are obtained from near‑real‑time user logs; however, direct use of historical logs has accuracy and flexibility issues. Online experiments randomly replace displayed items with rule‑generated candidates, sampling according to original rule trigger rates and adjusting sample weights to maintain proportional influence.

Evaluation focuses on click‑through rate and user churn, showing steady improvements since the feature launched in December 2015, with CTR rising from 10% (subjective Bayes) to 18% (RankBoost) and further gains after LambdaMart.

Future work includes adding finer‑grained rules, further tuning the LambdaMart model, and introducing LSTM‑based sequential models to predict users' next destination and demand from their overall behavior timeline.