Personalized Marketing Platform for Travel Scenarios: Architecture, Algorithms, and Cold‑Start Solutions at Alibaba Fliggy

Introduction The Fliggy personalized marketing team reveals how an "autonomous" platform supports thousands of travel scenarios with individualized recommendations, addressing challenges such as sparse user behavior, cold‑start during large promotions, and diverse product dependencies.

01 Background Fliggy serves both daily and event (e.g., Double‑11, 618) travel pages, delivering personalized placements across home, channel, and event pages.

02 Platform Architecture The architecture consists of four abstraction layers:

Scene abstraction (e.g., entry, single‑tab, multi‑tab, capsule, theme list, single/multi‑material, LBS recommendation, real‑time hot list).

Function abstraction (recall, ranking, weighting, scattering, slotting).

Link abstraction (product operation side, selection platform, scene management, personalized delivery, front‑end, user).

Personalized delivery (offline modeling of user/item, online real‑time context, matching, ranking, traffic control).

03 Platform Algorithms

3.1 User Session Understanding Users exhibit interests from basic attributes, group attributes, real‑time, periodic, and long‑term preferences. Sessions consist of actions such as homepage, search, product page, favorite, add‑to‑cart, and purchase. A heterogeneous graph is built where users, items, and POIs are nodes, and interactions form edges.

GraphSAGE is used to learn node embeddings (MRR ≈ 0.96, loss < 0.15). Session representation combines long‑term attention and short‑term recent item embeddings, followed by a fully‑connected layer.

3.2 Cold‑Start Techniques

Hierarchical topic‑based recall (U2Htopic2I) leveraging active user behavior with time decay to build a topic preference library.

Attribute‑to‑topic mapping (Attr2Htopic2I) learns user attribute → topic preferences, improving cold‑start recall.

Cross‑domain mapping (Taobao → Fliggy) using shared users to train an embedding mapper, enabling cold‑start users on Taobao to be served on Fliggy.

Heterogeneous relation modeling (HERS) that incorporates user‑user, user‑item, and item‑item relations for sparse data.

3.3 Travel Play‑Tag System Four steps are described:

Tag extraction using TF‑IDF, TextRank, and TextCNN.

Construction of a hierarchical tag tree covering 16 top‑level categories (food, culture, transport, accommodation, etc.) and detailed leaf tags (surfing, skiing, …).

Tag attachment to items via multi‑label classification, improved with LEAM embeddings.

Applications: tag‑based user‑item matching, scenario‑driven marketing, and time‑space‑aware recommendation.

Overall Solution The end‑to‑end pipeline integrates data backend (user/item layers, tag systems), a personalized marketing middle‑platform (algorithmic capabilities, full‑stack traffic control), and business front‑ends (operational platforms, marketing delivery, guided shopping).

References

Zhao et al., "IntentGC", KDD 2019.

Hamilton et al., "Inductive Representation Learning on Large Graphs", NeurIPS 2017.

Lv et al., "Neighborhood‑Enhanced and Time‑Aware Model for Session‑based Recommendation", arXiv 2019.

Wang et al., "Cross‑domain Recommendation for Cold‑Start Users via Neighborhood Based Feature Mapping", DASFAA 2018.

Hu et al., "HERS: Modeling Influential Contexts with Heterogeneous Relations", AAAI 2019.

Wang et al., "Joint Embedding of Words and Labels for Text Classification", arXiv 2018.