Interactive Recommendation and Travel Theme Recommendation in the Fliggy App

In Fliggy’s recommendation scenario, beyond recommending single items, the system also recommends collections of items organized as travel themes, which serve both precise matching and discovery of user needs.

Interactive recommendation adopts an explicit "question‑feedback" loop, allowing users to actively participate and enabling the system to capture real‑time preferences, which is especially useful in the waterfall‑style feed of the Fliggy app.

Compared with conventional Top‑N recommendation, interactive recommendation transmits click data through a high‑speed channel, updating user‑preference models instantly and providing a lightweight, locally optimal, dialog‑style experience.

User demands are categorized into decision‑phase, sprint‑phase, and seed‑phase users; each group receives tailored material such as similar‑item recalls, complementary‑item recalls, or exploratory queries.

Real‑time interest capture works by inserting personalized cards (e.g., destination cards) into the feed; user clicks reinforce the associated triggers, while non‑clicks act as negative feedback.

The recall stage includes three types: (1) similar‑item recall based on high relevance and destination constraints, (2) complementary‑item recall that matches hotels, tickets, or activities to a selected travel package, and (3) exploratory recall that expands query‑to‑item relationships to increase coverage.

Recommendation reasons are generated from user behavior, user‑group attributes, or theme‑ranking information and displayed on fixed card slots.

Card selection uses a nine‑type card pool; candidate cards are ranked by a three‑layer fully‑connected network using user profile, real‑time trigger, and card‑specific features, then the top‑1 is shown.

Travel‑theme recommendation follows a similar recall‑and‑ranking pipeline: first, personalized first‑image and copy are selected; then themes are recalled based on user status (silent, seed, decision, in‑trip) using location, behavior, and X2I/I2I channels; finally, theme ranking combines product CTR and theme CTR via multi‑task learning, sharing hidden layers while keeping separate output heads.

Engineering tricks such as up‑sampling positive samples from theme landing‑page clicks further boost CTR, and modeling the product‑theme pair directly simplifies the ranking process.

The overall system demonstrates how real‑time interactive recommendation and personalized travel‑theme recommendation improve user engagement and mitigate the “rich‑get‑richer” effect in large‑scale e‑commerce platforms.