Splicing Recall for Flight Ticket Search: Challenges, Algorithmic Solutions, and Online Impact

The talk introduces the concept of splicing recall in flight ticket search, a scenario where direct itineraries are unavailable or unsatisfactory, requiring the platform to combine multiple flight segments or even different transport modes to fulfill user requests.

Background and challenges are described: Fliggy's ticket search faces massive flight and merchant data, multiple search types (single‑trip, round‑trip, multi‑city), and constraints such as real‑time price volatility, complex business rules, and cross‑border visa requirements.

Three recall types are defined: single‑ticket recall, splicing ticket recall (combining tickets from different merchants without pre‑packaging), and cross‑category splicing (e.g., combining train and flight tickets). The focus is on the second type.

Key challenges include the diversity of business scenarios, rapid price changes that prevent exact price retrieval at every stage, and intricate merchant‑specific splicing rules.

The desired goals of splicing recall are to provide results for cold‑start routes, offer cheaper or more convenient itineraries, and increase user choice, especially during periods like the pandemic when direct flights are scarce.

Algorithmic practice is presented as a four‑step pipeline: (1) build a city connectivity graph using flight schedules and historical transaction data; (2) preliminarily select transfer cities based on low‑price offline caches; (3) refine transfer city candidates using a simple neural‑network model that scores static user features, flight‑combination features, and statistical attributes; (4) select flight combinations by ranking with a logistic‑regression model that incorporates price (using yesterday’s lowest price as a proxy) and travel time features.

During transfer‑city refinement, the NN model scores candidates using user static features, flight‑combination characteristics, and industry‑specific signals, selecting the top transfer cities.

In the flight‑combination stage, the LR model evaluates price (normalized by recent historical minima) and total travel time (using transfer‑waiting time as a proxy), along with feedback‑derived and convenience features, to produce the final ranked itineraries.

Online results show a noticeable increase in spliced ticket volume and a modest uplift in overall conversion rate. Case studies illustrate that the system can produce shorter, cheaper itineraries by selecting optimal transfer points and avoiding unnecessary airport changes.

The summary reflects on architectural insights (data layer, algorithm layer, application layer, and feedback loop) and shares practical experience: deep business understanding and feature engineering often outweigh complex models, thorough data analysis is essential for identifying optimization points, and close collaboration with business stakeholders drives effective algorithmic improvements.

Overall, the splicing recall framework enhances flight search robustness, expands user options, and improves conversion, especially under constrained flight availability.