How Tencent Ads’ CONFLUX and MVKE Algorithms Boost Conversion – Insights from KDD2022

CONFLUX Algorithm

The CONFLUX algorithm jointly optimizes contract (brand) and performance (conversion) advertising rankings to maximize overall system revenue. It treats the two ad types as complementary objectives and learns a unified scoring function that balances long‑term brand exposure with short‑term conversion likelihood. The model is trained on billions of historical ad impressions, using a multi‑task loss that combines CTR/CVR prediction for performance ads with a brand‑impact proxy for contract ads. To meet the latency constraints of a production ad serving system, the authors apply model distillation: a large teacher model is compressed into a lightweight student model that retains predictive power while reducing inference cost for hundreds of billions of daily requests.

Deployment scale: The algorithm is deployed in Tencent’s real‑time ad ranking pipeline, influencing tens of billions of ad impressions per day.

Validation methodology: Offline evaluation on held‑out logs (AUC, calibration) is followed by online A/B tests measuring revenue lift, click‑through rate, and brand‑awareness metrics.

Resource efficiency: Distillation reduces model size by up to 70% and cuts GPU/CPU inference time, enabling the algorithm to serve at the required sub‑millisecond latency.

Future directions: Extending brand‑ad post‑click effectiveness modeling, mining optimal placement strategies from large‑scale data, and aligning performance‑ad objectives with downstream business metrics such as order count and payment amount.

Full paper (Chinese):

http://mp.weixin.qq.com/s?__biz=MzIzMzgzOTUxNA==&mid=2247487819&idx=1&sn=6ad8d6b68676955e3b7bad657286eb60#wechat_redirect

MVKE Model

MVKE (Multi‑Goal User‑profile Modeling) is a high‑throughput framework that simultaneously learns multiple user‑profile objectives (e.g., click, conversion, dwell time) to improve recommendation efficiency across heterogeneous scenarios, including advertising, feed, and search. The architecture consists of:

A shared embedding layer that encodes user and item features.

Separate task‑specific heads that predict each objective using a shared backbone.

A multi‑task loss that balances gradients via dynamic weighting, allowing the model to prioritize scarce conversion signals while still learning from abundant click data.

Key technical aspects:

Scalability: Designed for billions of inference requests per day; supports model parallelism and mixed‑precision training.

Feature enrichment: Item representations can be extended with rich metadata (category, textual description, visual embeddings) without changing the core architecture.

Adaptability: To apply MVKE to a new domain, replace or augment the item feature set and define task‑specific target labels; the shared backbone remains unchanged.

Evaluation: Offline metrics (multi‑task AUC, precision@k) and online A/B tests show reduced latency and higher overall conversion rates compared with single‑task baselines.

Full paper (Chinese):

http://mp.weixin.qq.com/s?__biz=MzIzMzgzOTUxNA==&mid=2247487818&idx=1&sn=6cef12744347e542898a9f8611e54ea0#wechat_redirect