Ctrip CTO Gan Quan on Building a Data‑Driven Personalized Recommendation System

In the era of consumption upgrade, time has become a new battlefield and currency. Travel platforms like Ctrip aim to win users by providing personalized services that optimize users' time rather than creating addiction.

Since 2015, Ctrip has used cross‑screen data processing, real‑time APIs, and data‑model training platforms to predict user behavior and deliver personalized pre‑travel recommendations on web pages and EDM, boosting click‑through and conversion rates.

Over the past two years, Ctrip’s personalization has expanded beyond pre‑travel suggestions to real‑time, cross‑device advertising and product recommendations based on user searches, abandoned hotel bookings, and other interactions.

CTO Gan Quan emphasizes that the priority is to show users what they are interested in, making the search and booking process faster and more convenient.

A Hundred Million Small Goal, Personalized Recommendation to Achieve

From 2015 to present, Ctrip has continuously improved its big‑data platform while expanding user scenarios and deploying richer recommendation models.

The platform integrates data from almost all internal business lines and many external third‑party sources, forming a rich tourism knowledge base and data center.

Algorithmically, Ctrip has broadened its deep‑learning toolbox, selecting and iterating models based on user and data performance, and has connected the recommendation engine to over 60 business lines and external advertising channels.

The recommendation engine enables cross‑domain personalization, such as suggesting hotels, car rentals, tours, and shopping items after a user searches for a flight.

Most Ctrip app channels—homepage recommendations, search, discount flight listings—use the personalized recommendation platform, with A/B testing guiding the optimal matching between users and products.

Through A/B testing, both column‑level personalization and overall site stickiness have shown significant improvement, leading to notable revenue gains.

Ctrip’s small data team (under 20 people) handles data collection, cleaning, storage, and modeling, and operates the personalized recommendation platform that has driven a 1.2 billion CNY revenue increment in 2016 with a high ROI.

In a specific business scenario, integrating the recommendation platform with only 1–2 engineers increased that line’s revenue by 6% in 2016.

Personalized recommendation remains a challenging problem for OTA platforms, especially when spanning multiple business lines, and there are few industry benchmarks to follow.

Personalized Recommendation Loop Emphasizes Speed

Ctrip collects user data from all terminals—mobile app, H5, web, WeChat mini‑programs—capturing searches, browsing, orders, and after‑sale service across devices in real time.

Traditionally, data collection required pre‑defined “tracking points” (埋点) and client updates for new dimensions, a lengthy process.

Ctrip now uses a self‑developed, non‑intrusive User Behavior Tracking (UBT) system that captures all user actions in real time without prior instrumentation, greatly improving data acquisition efficiency.

External factors such as weather, water temperature, and wind also influence travel decisions, so Ctrip collaborates with third‑party providers to enrich its recommendation inputs.

Data from heterogeneous external sources must be normalized into a standard structure suitable for Ctrip’s systems.

The integrated data feeds not only individual personalization but also influences recommendations for other users, achieving a second‑level data loop where user actions instantly affect the algorithm’s output for the broader audience.

For example, a surge in searches for spring outings like “flower‑viewing” triggers the platform to identify hot content and proactively push it to similar users based on location, history, and other dimensions.

Gan Quan summarizes that the end‑to‑end flow—data collection, cleaning, storage, modeling, and real‑time application—forms a closed loop that drives personalized recommendation and marketing.

Personalization Relies on Data Quality

Travel differs from traditional e‑commerce: users seek novelty rather than repeat purchases, making inspirational, personalized recommendations crucial.

Ctrip has experimented with scenarios such as pre‑booking popular hotels, chartering flights for hot routes, and predicting emerging destinations, with deeper, finer‑grained use cases still to be explored.

Factors influencing travel decisions include trip length, personal preferences, destination knowledge, cost, resources, weather, and safety, leading to long decision cycles (average 20 days for domestic trips, over 40 days for outbound trips).

The “Discover” channel, an AI‑driven personalized destination recommendation system, aggregates user interest points, historical data, and contextual variables (price, season, weather, shopping, food, family, couple, etc.) to match users with global destinations and comprehensive itineraries.

Unlike traditional travel guides that require users to read extensive content and verify safety, the Discover channel streamlines decision‑making by presenting tailored options, reducing friction.

Since its 2017 pilot, the Discover channel has been tested with 10 % of users, showing better performance than simple personalization in A/B tests, and plans to incorporate user reviews and community data in the future.

Collaboration Between Technical and Business Teams

Successful personalization depends on close cooperation between technology and business units, sharing metrics and aligning goals.

Ctrip’s data‑driven successes have encouraged business teams to embrace data collaboration, while engineering passion is fostered through mechanisms that recognize impactful contributions.

To resolve disputes, Ctrip established a Big Data Committee in late 2016, providing a forum for discussion and consensus before implementation.

Gan Quan emphasizes that long‑term progress in big data, AI, and personalization requires problem‑driven iteration, micro‑adjustments, and continuous validation.

Afterword

Three essential elements for effective personalization are: massive data accumulation, adaptable system architecture and algorithms for diverse scenarios, and concrete business applications that validate user acceptance.

Ctrip’s strengths lie in traffic volume, data richness, and a wide array of business lines, but the ultimate advantage comes from efficient system and algorithm design.

While the era of “land‑grab” online travel is over, the new competition focuses on fine‑grained operations, precise user matching, and data‑driven incremental growth.