Construction and Application of a User Profile Tag System: Methods, Platforms, and Use Cases

The talk introduces the concept of a profile tag system, explaining why Qunar needed to integrate multiple business‑specific tag schemas into a unified framework to support marketing, risk control, and business analysis.

It outlines the five main parts of the presentation: the tag taxonomy, the tag platform, common algorithmic tags, the construction process, and application scenarios, followed by a Q&A session.

Tag Taxonomy : Tags are categorized by business needs (marketing, risk control, internal analysis, user description) and by construction method (statistical, rule‑based, model‑based). The speaker emphasizes the importance of consistent definitions across business units and the challenges of merging divergent tag definitions.

Tag Platform : The CDP platform at Qunar handles tag generation, storage, and serving. It supports both offline and real‑time tags, with update frequencies ranging from hourly to real‑time streaming, and offers flexible access via Redis, HBase, or other storage back‑ends depending on latency requirements.

Construction Methods : Statistical tags rely on simple SQL queries, rule‑based tags are crafted by analysts familiar with business logic, and model‑based tags require data‑science teams to train algorithms, often facing data sparsity and accuracy challenges.

Update Cycle & Access : Besides scheduled hourly, daily, and monthly updates, the system implements real‑time streaming updates using technologies such as Flink or Spark, allowing complex tag logic to be expressed via uploaded Python or SQL code.

Algorithmic Tags : Common model‑based tags include classification, recommendation, knowledge‑graph, causal inference, image processing, NLP bots, and look‑alike algorithms. The speaker discusses how each algorithm type is applied to user or item profiling.

Application Scenarios : The tags are used for marketing audience selection and expansion, business metric attribution analysis, and A/B experiment effectiveness analysis. Each scenario is illustrated with workflow steps and the role of tags in improving decision‑making.

Q&A Highlights : Questions cover the distinction between user behavior and business logs, implementation of streaming tags, definition of real‑time tags, ID mapping strategies, product‑level tag use cases, the need for custom development of real‑time tags, tag lifecycle management, sample size estimation for experiments, storage of multi‑label user profiles, and the involvement of profiling engineers in recommendation pipelines.

The session concludes with a summary of the key takeaways and thanks to the audience.