Travel Demand Prediction and Recommendation Optimization at Fliggy: Challenges, Algorithm Evolution, and Future Directions

The presentation introduces the background and challenges of travel‑related services (air tickets, train tickets, car tickets) on the Fliggy platform, highlighting four user travel stages (demand ignition, pre‑travel, in‑travel, post‑travel) and four key difficulties: long decision cycles, sparse behavior, temporal sequencing, and spatiotemporal correlation.

Algorithm optimization is described in two phases: recall and ranking. The recall phase improves traditional Swing methods by capturing user intent (e.g., H2S, S2S, W2W) and using heterogeneous graphs that combine spatial and popularity information. The ranking phase introduces single‑task models (AutoInt, WDL, DIN) and multi‑task models (ESMM, MMOE, PLE), followed by advanced modules such as LSGMNet, G‑PDIN, and intention‑capture mechanisms.

Key technical components include:

Construction of spatial and hot‑city graphs processed by GraphSAGE.

Intention capture using capsule networks and attention.

Periodic behavior modeling with dilated convolutions.

Multi‑task learning with GeoHash and city‑level sub‑tasks.

Offline experiments on Fliggy and Fliggy2 datasets show AUC improvements from 0.80 to 0.83, while online A/B tests report ~1% gains in PVCTR and UVCTR after deploying the new models.

Future work focuses on ultra‑long behavior sequence modeling, spatiotemporal graph construction, meta‑learning, and cross‑domain transfer to address cold‑start issues across different travel modes.

The Q&A section addresses the publication status of G‑PDIN, features used in GAT nodes, the impact of popular route recommendations, and details of long‑sequence modeling and time‑information handling.