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

The Fliggy team introduces a "driverless" personalized marketing platform that automatically supports thousands of travel scenarios with individualized recommendations for both daily and promotional events.

Background : Travel pages contain diverse entry points, tabs, capsules, and recommendation modules, each requiring tailored material such as products, hotels, tickets, or POIs. The platform must handle a large variety of pages, heterogeneous data sources, and frequent operator interventions while maintaining a unified model for efficiency.

Platform Architecture : The design consists of four abstractions – scene abstraction (e.g., entry, single‑tab, multi‑tab, capsule, theme list, single/multi‑material, LBS recommendation), function abstraction (recall, ranking, weighting, dispersion, fixed‑position), link abstraction (product runtime, selection platform, scene management, personalization engine, front‑end, user) and a full‑stack flow that integrates data backend, personalization middle‑platform, and business front‑end.

User Modeling : A four‑layer pipeline (real‑time public, real‑time feature, user expression, real‑time service) captures static attributes, group traits, real‑time intent, and long‑term preferences, exposing services such as user state, intent vectors, and profile tags.

Full‑Domain Traffic Control : Real‑time stream computation, configurable product control, and a PID‑based control center adjust traffic for hot items, new products, and promotional goals, with results visualized on a live dashboard.

Recommendation Algorithms : The system combines offline user‑item graph modeling (GraphSAGE) with online session understanding (attention‑based long‑term and short‑term fusion), neighbor session retrieval, and co‑attention mechanisms, achieving high offline metrics (MRR≈0.96, HIT@20≈0.6 for items).

Cold‑Start Techniques : Two approaches are described – hierarchical topic‑based recall (U2Htopic2I) that maps user attributes to topic preferences, and attribute‑to‑topic mapping (Attr2Htopic2I) that learns the relationship via a neural model, both yielding multi‑point CTR lifts.

Cross‑Domain Mapping : User embeddings from Taobao are transferred to Fliggy using a two‑layer fully connected network that incorporates behavior sequences, LBS, and attribute information, improving cold‑start recall by 1‑2 % on Fliggy and 2 % on Taobao.

Heterogeneous Relation Modeling : User‑user, user‑item, and item‑item relations are encoded in a heterogeneous graph, with user groups and item groups constructed from recent behavior to enhance similarity scoring for cold‑start users.

Travel Play‑Tag System : A four‑stage pipeline (keyword extraction via TF‑IDF, TextRank, TextCNN, label embedding with LEAM) builds a hierarchical tag tree covering 16 top‑level categories (e.g., food, culture, transport) and fine‑grained leaf tags (e.g., surfing, skiing). Tags are attached to items through multi‑label classification, enabling tag‑based search, scenario‑driven delivery, and time‑space‑aware recommendations.

The article concludes with references to related research on heterogeneous graph recommendation, inductive graph learning, session‑based models, cross‑domain mapping, and joint word‑label embedding.