Interactive Recommendation and Travel Theme Recommendation in the Fliggy App

The presentation introduces two sub‑topics: practical applications of interactive recommendation and deep‑learning algorithms for travel‑theme scenarios in the Fliggy app.

Interactive recommendation, first proposed by YouTube in 2018, uses a question‑and‑feedback loop to explicitly capture user preferences, enabling real‑time preference updates in the app's waterfall feed.

Compared with traditional top‑N feeds, interactive recommendation provides explicit user interaction, faster data pipelines, and higher local CTR, offering a more engaging experience.

User demand is categorized into decision‑phase, sprint‑phase, and seed‑phase users, each with tailored recommendation strategies such as similar‑item recall, complementary recall, and exploratory recall.

The system employs three types of recall: similarity‑based, relevance‑based, and exploratory, leveraging queries, destination cards, and multi‑hop item associations to broaden coverage.

Recommendation reasons are generated from user behavior, user group attributes, and item ranking, displayed on fixed card slots.

Card selection uses a three‑layer fully connected network with features from user profile, real‑time triggers, and card attributes, achieving higher CTR than conventional waterfall recommendations.

Travel‑theme recommendation involves personalized image and copy selection, multi‑stage recall (item and theme), and ranking that jointly considers item and theme CTR predictions, using multi‑task learning to share representations.

Various user segments (silent, seed, decision, in‑trip) receive customized theme recalls based on location, intent, and behavior.

Engineering tricks such as up‑sampling positive samples from theme landing pages further improve performance.

The talk concludes with thanks and invites the audience to share, like, and follow.