How Youku Scaled Millions of DAU with DSP‑Driven RTB Advertising Algorithms

1. User‑Growth Business Background

Since 2019, Youku has used DSPs on platforms such as Toutiao and Alibaba Mama to deliver video ads, achieving stable user growth and eventually attracting millions of daily active users (DAU). The goal is to maximize high‑value DAU—new, low‑activity, and silent users—while keeping costs under control.

2. RTB Overview

Real‑time bidding (RTB) connects the DSP (advertiser’s bidding agent) with an ad exchange (ADX) such as Toutiao ADX or Alibaba TanX. When a user visits a web page or app, the request goes to the ADX, which forwards it to multiple DSPs. Each DSP uses user features from a DMP to estimate value and returns a bid within 100 ms. The ADX selects a winning ad based on ranking and pricing rules, then reports the click back to the winning DSP.

Key RTB components:

Bid mechanism: Youku uses GSP (second‑price) with CPM pricing.

Business‑specific constraints: CPM requires internal value modeling using internal data; cost control varies across user states; a single‑user DSP must integrate multiple ADX platforms.

3. Advertising Ranking Algorithm and Optimization

3.1 Algorithm Flow

The offline pipeline generates millions of candidate ads, then:

Filter low‑quality ads (e.g., long‑standing or low‑CTR ads) via manual or algorithmic selection.

Perform multi‑channel personalized recall to retrieve hundreds of ads per user.

Apply rule‑based filters (duplicate removal, exposure frequency caps).

Score ads with a prediction model, compute a bid price, and return the ad and price to the ADX.

The process borrows from recommendation systems but adds CTR bias correction and a pricing step after estimation.

3.2 Implementation Choices

Challenges include a strict 100 ms response time, multi‑channel integration, and high traffic volume. Common ranking models are:

Logistic Regression (LR) – requires extensive feature engineering.

Gradient Boosted Decision Trees (GBDT) – fast training/inference, but can overfit categorical features.

Deep Neural Networks (DNN) such as DCN, WDL, MMOE – handle diverse features via embeddings but consume large memory and have higher latency.

To unify models across channels, platform identifiers are used as features, and statistical aggregates per platform are added, enabling a transition from single‑channel LR models to a unified GBDT model with measurable business gains.

3.3 Problems and Reflections

Traditional item‑based recall assumes users like what they watch, leading to three issues:

Viewed videos may not reflect true preference.

Similarity of videos does not guarantee relevance to the user’s interest.

Low‑frequency users have very few views, reducing recommendation diversity.

From a causal inference perspective, silence arises from unsatisfactory recommendations, while retention stems from liked content. The team built an unbiased similarity metric, matched low‑activity users to high‑activity “counterfactual twins,” and delivered tailored recommendations, achieving significant uplift.

4. Automated Bidding Algorithms

4.1 Formal Problem Definition

Advertising bidding faces different constraints for daily and burst (high‑volume) campaigns. Daily campaigns have budget, first‑activation cost, and user‑value constraints; burst campaigns have ample budget but still respect first‑activation cost and user‑value constraints.

4.2 Single‑PID Solution for Burst Campaigns

Under burst conditions, the optimal bid equals estimated CTR/CVR multiplied by a controllable coefficient α. A single controller adjusts α to satisfy cost limits while improving conversion.

Controllers compute a signal from feedback data using a control function; the signal feeds an actuator that exponentiates the value to produce the final bid. Common controllers include PID (proportional‑integral‑derivative) and wave‑based controllers, with PID offering finer adjustment.

4.3 Daily Campaign Solution

Daily campaigns are modeled as a multi‑objective constrained maximization problem solved via Lagrangian duality, yielding a bid formula with two parameters p and q. Two PID controllers regulate p and q; the q controller includes a correction signal targeting final CPC, and a linear interpolation models the coupling between the two PIDs.

5. Bidding System Engineering

The request flow through the DSP management backend proceeds as follows:

Logs of bidding, impressions, and clicks are collected, parsed by Blink, and stored in an in‑memory database along with historical error data.

Every M minutes, an offline job generates a delivery plan (cost, budget, algorithm parameters). The bidding core reads error data, computes current bid parameters, and writes them back to the in‑memory store.

The TPP platform performs ad recall and CTR estimation, then reads the bid parameters to return the ad and its price.

6. Summary and Future Plans

The practice required balancing theoretical algorithmic choices with practical constraints such as machine resources and manpower. By layering the problem, applying A/B testing, and tailoring models to business specifics, the team demonstrated how algorithmic engineering can serve user‑growth goals.

Future directions include extending CTR estimation to user‑value prediction, exploring reinforcement‑learning‑based bidding, and expanding the framework to acquisition, hot‑show promotion, and membership growth.