Real-Time User Understanding Service (RTUS) for Travel: Architecture, Algorithms, and Experimental Evaluation

The article introduces RTUS, a real‑time user understanding service for the Fliggy travel scenario, outlining its two main parts: the service architecture and algorithmic practices.

It first describes travel‑industry behavior characteristics—low frequency, spatio‑temporal attributes, and periodicity—and how user actions in Fliggy follow three typical chains (search‑driven, navigation‑driven, and browsing‑driven). Based on these traits, a unified real‑time feature service (RTFS) was built, collecting logs from multiple terminals and scenarios and defining a common user‑behavior schema.

RTUS evolves from RTFS by aggregating data from different chains (server logs, client logs, special‑chain logs) into a standardized schema, enabling asynchronous fusion that provides accurate, low‑latency user‑behavior data for downstream services.

The service architecture consists of three layers: a real‑time public layer (raw logs), a feature‑processing layer (standardization, aggregation, and generation of LBS, sequence, statistical, and item‑feature vectors), and a user‑expression layer where various models predict short‑ and long‑term preferences, real‑time intents (e.g., POI, price), and travel‑state representations. Model outputs are stored in a real‑time graph database for downstream queries.

Algorithmic practice focuses on three challenges:

Cold‑start user interest representation (SMINet): multi‑aspect interest extraction using query‑driven top‑k item sets for spatio‑temporal, group, periodic, long‑term, and short‑term interests, combined with a time‑aware GRU and co‑attention mechanisms.

User travel‑intent prediction (DCIEN): a dual‑channel network that fuses online behavior sequences with offline holiday/weekly patterns via periodic‑aware attention and target‑aware attention, achieving superior Recall and MRR compared with DSIN, DMSN, etc.

User destination prediction (OOPIN): offline‑online fusion of city‑visit matrices and online behavior sequences, employing distance‑aware attention and information‑gain networks to predict future travel destinations, outperforming baseline methods.

Extensive offline experiments on Fliggy and public datasets, as well as online A/B tests, demonstrate significant improvements in recall, MRR, and CTR, especially for cold‑start users. Ablation studies confirm the contribution of each component (short‑term interest, spatio‑temporal pattern, offline‑online fusion, etc.).

The article concludes with a Q&A session covering implementation details such as query‑based interest extraction, evaluation metrics, data deduplication across multiple logging chains, and the generation of spatio‑temporal interest keys (e.g., Wuhan‑March).