Hubble Intelligent Audience Platform: Three‑Generation Algorithm Evolution for Mobile Marketing

Facing increasingly competitive mobile financial service scenarios, the core challenge for marketing operations is how to quickly and accurately reach the right users with suitable products and services. Ant Group’s Hubble Intelligent Audience Platform addresses this by providing a system that identifies high‑value users for each business scenario, improving traffic efficiency.

The platform comprises six major functions: tag‑based crowd creation, algorithmic crowd creation, real‑time tagging, crowd management, crowd insight, and effect analysis. Its algorithmic core has undergone three generations, each published in top conferences (KDD 2020, CIKM 2020).

Problem Definition : Given a marketing campaign, seed users, and a candidate pool, the goal is to select an audience that maximizes click‑through and conversion rates—known as audience expansion or look‑alike modeling.

First Generation : An end‑to‑end model based on Microsoft’s Deep Structured Semantic Model (DSSM). It splits user and campaign features, learns a joint embedding, and ranks candidates by preference scores. While it improves click‑through and conversion rates over traditional GBDT, it suffers from long training time and high resource consumption.

Second Generation : Introduces an asynchronous architecture separating offline graph neural network (AD‑GNN) and online lightweight learning (KD‑AE). The offline AD‑GNN builds a user‑campaign bipartite graph, learns embeddings via link prediction, and the online KD‑AE model, instantiated per request, performs knowledge‑distilled audience expansion using both hard labels (seed vs. non‑seed) and soft labels from the teacher AD‑GNN.

The offline AD‑GNN model consists of three modules:

Projection : Maps each node’s raw features into K sub‑spaces representing different user intents.

Neighborhood Routing : Iteratively aggregates neighbor information to produce decoupled user embeddings.

Adaptive Breadth : Introduces an adaptive breadth score to down‑weight noisy edges, forming an adaptive & disentangled layer that can be stacked L times.

Online KD‑AE Model : Uses the user embeddings from AD‑GNN as teacher knowledge. It trains a lightweight student model with a loss that combines hard labels and softened teacher predictions, enabling fast inference on new campaigns.

Third Generation : Adds incremental learning to the asynchronous GNN framework. Real‑time feedback (clicks vs. exposures) collected during a campaign is used to continuously update the model. A meta‑learner learns a weighting function for seed users, mitigating coverage bias when seed quality varies across campaigns.

The overall system thus combines pre‑campaign historical data, real‑time feedback, and meta‑learning to deliver efficient, high‑quality audience expansion for mobile marketing.

References: deWet & Ou 2019; Ma et al. 2016; Liu et al. 2019; Huang et al. 2013; Hinton et al. 2015; Shu et al. 2019.