Evolution of Display Advertising Effect Optimization at 360: System Architecture, Smart Bidding, and Model Advances

Guest and Context Speaker: Liu Bin, Senior Algorithm Engineer at 360. Editor: Zhao Yong. Produced by DataFunTalk.

Background Introduction Display advertising involves media pages (e.g., PC portals, mobile apps) sending user flow information to an Ad Exchange, which forwards the request to multiple DSPs. DSPs select suitable ads based on targeting criteria, and the Exchange returns the highest‑bidding ad to the media within a 100 ms response window.

Common Display Ad Types - Scene window ads (shopping tab with product name and price) - Splash screen ads (full‑screen creative at app launch) - Feed ads (native‑style ads within content streams).

Display Advertising Architecture The architecture consists of online and offline systems. Online, the DSP server receives traffic, forwards it to Ad Search for coarse recall, then to Ad Selector for final selection; all events are logged to Kafka and stored in HDFS. Offline, ETL processes generate training samples for CVR/CTR models, which are pushed online for use in recall and ranking.

CPC Advertising Settings CPC ads charge advertisers per click. The typical CPC setup includes uploading creative assets, selecting traffic filters (user, media, time, location), and setting the bid price per click.

Effect Optimization Overview Conversion is defined as any post‑click action valuable to the advertiser. Effect optimization aims to increase conversions while reducing cost per conversion. 360’s philosophy is that every option can be optimized, dividing optimization into advertiser‑initiated and backend‑automated strategies.

Traffic Layering Traffic is visualized as a pyramid: low‑quality traffic at the base receives budget caps or discounted bids, while high‑quality traffic (selected via conversion‑rate models) receives higher bids.

Smart Click Bidding Smart bidding adjusts CPC based on real‑time feedback and conversion‑rate models, applying premiums to high‑conversion segments and discounts to low‑conversion segments.

Intelligent Creative Combination Advertisers upload multiple images and copy texts; an algorithm selects the best image‑copy pair for each impression, automatically closing under‑performing combinations.

Creative Combination Generation Network A reinforcement‑learning‑based generator creates image‑copy pairs, scored by a conversion‑rate prediction model that uses attention over image, text, and context features.

Ad‑Delivery Intelligence Evolution 1. oCPC (optimized CPC) simplifies the UI, quantifies advertiser goals (conversion type and target CPA), and balances efficiency with volume. 2. Bidding logic evolves from CPC × PCTR to a two‑stage approach: early stage uses system‑provided CPC, later stage uses CPA × PCVR × PCTR. 3. Model strategy shifts from multiple per‑customer models to a single unified model as data scales. 4. Model complexity progresses from simple GBDT to LR, then to DNN as feature volume grows. 5. The oCPC model adopts a PNN architecture with grouped embeddings (ad, media, user, context) and optimized online inference. 6. Task design moves from single‑task conversion prediction to multi‑task (CTR, CVR, etc.) sharing embeddings. 7. Bidding expands from single‑goal to multi‑goal (e.g., registration and payment) using MMOE experts. 8. Latency is reduced via knowledge distillation, improving the teacher model and adding pre‑conversion features.

Conclusion The presentation summarized the full lifecycle of display ad effect optimization at 360, illustrating how systematic architecture, data‑driven models, and intelligent product design have driven the evolution toward fully automated, low‑latency, multi‑objective advertising solutions.