Machine Learning Applications in OTA Hotel Industry: From Data Challenges to Value Creation

Pan Pengju, a BI manager from Ctrip hotel R&D, introduces the use of machine learning and large‑scale data to overcome challenges in the online travel agency (OTA) hotel sector and enhance the booking experience.

The OTA hotel industry differs from other sectors due to limited, time‑bound inventory, reliance on third‑party agents, and self‑operated rooms, leading to challenges such as over‑selling, agent coordination, and service KPI issues like "no‑room" and "no‑order" incidents.

Project benefit evaluation is performed by quantifying direct and indirect customer costs (waiting time, loss per incident) and comparing them to the effort required for algorithmic improvements, turning customer experience into monetary value.

Algorithmic practice starts with massive data volumes (2 billion PV, 10 TB/day) split into business, performance, behavior, and crawler data; the talk emphasizes the importance of hot, timely data over cold historical data for real‑time impact.

Model evaluation methods covered include A/B testing, alternative offline metrics, and log‑based validation combined with risk control, highlighting the need for both accuracy and operational safety.

Concrete case studies are presented: (1) order‑volume forecasting using ARIMA with seasonal factors achieving ~5 % error; (2) confirmation‑time reduction by inserting a predictive model into the workflow, improving speed with 93 % accuracy; (3) inquiry‑room modeling to automate opening/closing of rooms based on predicted availability; (4) user price‑preference prediction using XGBoost, attaining 77 % accuracy within ¥50.

Experience sharing covers a full model lifecycle—data validation, feature engineering, handling missing values with auxiliary models, limited normalization for tree‑based methods, categorical processing, model fusion, and offline‑to‑online tuning to correct data errors.

The deployment architecture uses Java wrappers to serve models stored as R or Python files, invoking RServer/PyServer for predictions, with a comprehensive API and data‑validation checks to ensure reliable online operation.