Intelligent Customer Acquisition System Practice at Du Xiaoman Financial

In this talk, the speaker outlines the practice of Du Xiaoman Financial's intelligent customer acquisition system, summarizing the process into four stages: channels, efficiency, data understanding, and the acquisition platform itself.

Channels – The discussion starts by asking key questions about acquisition volume, target audience, available channels, and cost. Four main channels are identified: direct outreach using internal data, bidding advertising, traffic cooperation, and offline channels, each with distinct advantages and trade‑offs.

Choosing a channel depends on product maturity: early products rely on perception‑based channel selection, growing products use user‑profile analysis, and mature products shift to data‑driven model selection that can segment customers with high information utilization and robustness.

The efficiency part explains how a suite of models (response, pre‑approval, credit limit, price, loan, profit) improves per‑customer efficiency. By estimating credit cost and ranking customers, the system prioritizes low‑cost prospects for ad placement, while model filters remove low‑quality users in both proactive and passive outreach.

Multiple staged models are necessary because user churn differs across stages and the same features behave differently in various models, allowing selective feature usage.

Data Understanding – Data is described as the fuel for acquisition. Two aspects are highlighted: feature understanding (using DeepFM to capture both wide and deep interactions, achieving 1‑2% AUC improvement over XGBoost) and sample understanding (building a full‑lifecycle data pipeline to attribute conversions across multiple touchpoints such as feed ads, SMS, TikTok ads, app downloads, search ads, phone calls, etc.).

The platform architecture is then introduced, consisting of data, technology, model, system, function, and channel layers. The system layer includes an intelligent engine (XGBoost, DeepFM, stacking), a creative engine (GAN‑based text generation and image factor extraction), an ad‑delivery platform, a content‑assistance platform (NLP), and an experiment platform for rapid testing.

Finally, the article ends with a recruitment notice for senior machine‑learning engineers at Du Xiaoman Financial, outlining responsibilities such as building acquisition models, leveraging massive data, integrating marketing channels, and staying abreast of the latest AI research.