Technical Architecture and Strategies of Youku's Demand‑Side Platform (DSP)

With the rise of Real‑Time Bidding (RTB) online ad exchange models, many companies have built their own Demand‑Side Platforms (DSP) for user acquisition. Youku has also built a DSP in recent years and continuously iterated on it. This article shares the technical explorations, pitfalls, and solutions encountered during that process.

Business Goal : The core objective of the DSP is user growth—acquiring new users and re‑engaging existing ones to improve retention and activity. Users are segmented by monthly activity into new, low‑active, medium‑active, high‑active, and churned users. The focus is on recalling (re‑engaging) the new, low‑active, medium‑active, and churned users.

System Architecture : The DSP is organized around four capability dimensions: channel capability, strategy capability, algorithm capability, and real‑time monitoring & data.

1. Channel Capability

The system must integrate strong channel capabilities to reach a large user base. Only RTB‑related ad exchanges (e.g., Baidu ADX, NetEase, Guangdiantong, Tanx) are considered. Evaluation criteria include traffic volume, cost, and user‑segment distribution per channel. Key metrics include installation ratio, cross‑media overlap, reachable user count, click‑through and conversion rates, budget allocation, and optimal media mix.

2. Strategy Capability

Various delivery strategies are employed:

Frequency control (bidding, impression, click, and app‑launch frequencies) to avoid over‑exposure.

Material horse‑race (A/B testing of multiple creatives) to allocate more traffic to high‑CTR assets.

Cold‑start logic: initially distribute traffic equally among creatives for 15 min, then rank by CTR and allocate 80 % of traffic to the top creative, the remaining 20 % split among the others.

Budget smoothing to spend advertiser budgets evenly over time rather than exhausting them in the first minutes.

Retargeting (visitor‑recall ads) to bring back users who abandoned conversion.

Visual optimization of image creatives, including channel‑specific designs and AB testing of templates.

3. Algorithm Capability

Personalized off‑site ad recommendation is achieved through multiple recall channels:

Playback‑based recall (decay over time).

Hot‑program recall (e.g., newly released series).

User‑preference recall (stars, genres).

Behavioral recall (favorites, clicks, searches).

Channel‑specific models are trained because each ad exchange exhibits distinct feature distributions.

4. Data & Monitoring Capability

A comprehensive monitoring system is essential for optimizing the conversion funnel and reducing costs. The funnel stages include bidding participation, winning bids, impressions, clicks, app launches, playback, and next‑day retention. Each stage can be improved via:

Better user identification to increase bidding participation.

Higher bids and low‑latency responses (<100 ms) to improve win rates.

Creative and title optimization for impressions.

Accurate recommendation algorithms and appealing visuals for clicks.

Fast app launch and reduced pre‑roll length for higher launch rates.

Performance‑optimized playback pages for higher play rates.

Enhanced user experience for retention.

Full‑stack monitoring tracks metrics such as sudden spend spikes, CTR drops, and bidding volume declines to prevent budget waste.

Summary

Key considerations when building a DSP:

Channel integration and traffic acquisition.

Solid delivery strategies.

Algorithmic models for personalized recommendation.

Robust monitoring and data analysis.

Future work includes traffic value estimation and bidding price algorithms. The system has already proven valuable for user acquisition and retention, and the team welcomes experts in computational advertising to contribute.

Contact: [email protected]