Building User Interest Tags in WeChat's Recommendation System

Background: User profiling is crucial in recommendation systems, influencing recall and ranking. It includes basic user info and behavior-derived interest tags.

Problem Modeling: Directly feeding raw behavior to models is infeasible due to performance and cold-start issues. The authors frame user interest tag estimation as a multi-class classification task, placing tags on the label side to enable efficient training over a large tag space.

User-side Feature Learning: They propose a hierarchical attention mechanism for intra-domain and inter-domain feature fusion, optionally using multi-head attention to capture diverse interests, followed by a dense feature crossing module that computes pairwise interactions similar to FM, normalized and concatenated with linear features.

Click Tag Preference Modeling: Instead of treating all tags of a clicked article as positive examples, they aggregate tag preferences to model click probability, allowing the model to learn which tags are more influential, outperforming traditional split and attention‑based alternatives.

Evaluation Metrics: Online accuracy and coverage are measured by “有点数” (number of tagged impressions) and “有点率” (tag click‑through rate).

Experiments: Offline tests on internal and MovieLens data show that the proposed attention‑based feature fusion and feature crossing outperform baselines such as YouTube‑DNN, AFM, NFM, DCN, and AUTOINT. Joint loss training of multiple tags per article improves over split training. Online A/B tests confirm improvements in both tagged impression count and tag CTR.

Conclusion and Future Work: The authors highlight challenges in learning multi‑domain user features and better linking click behavior to tag, category, and public‑account preferences as directions for further research.