Design and Optimization of External Advertising Algorithms for User Growth at Youku

From 2019, Youku began using DSP to place video ads on platforms such as Toutiao and Alibaba Mama, combining user growth and advertising bidding. By borrowing practices from recommendation systems and leveraging its unique business context, a set of algorithms was developed to achieve stable user growth while keeping costs and budgets under control, ultimately delivering multi‑million DAU through effective traffic acquisition.

User Growth Business Background

The user growth process categorizes users into different states (new, low‑activity, medium‑activity, high‑activity, and silent). New, low‑activity, and silent users are valuable for increasing app activity. Real‑time bidding (RTB) is a primary method: DSP participates in ad auctions, winning ads are displayed on external platforms, and clicks redirect users to the Youku app, contributing to video consumption and subsequent visits.

The main business goal is to maximize high‑value DAU under controllable cost, attracting more new, low‑activity, and silent users.

RTB Overview

RTB consists of two key components: the DSP (advertiser’s bidding agent) and the ADX (auction platform such as Toutiao ADX or Alibaba TanX). The bidding flow involves the ADX forwarding user requests to multiple DSPs, each estimating value based on user features from a DMP and returning an ad and bid within 100 ms. The ADX selects an ad based on ranking and billing rules.

Key mechanisms include GFP (first‑price), GSP (second‑price), and VCG; billing models include CPM, CPC, CPA, and OCPX. Youku adopts GSP with CPM, which requires internal modeling to estimate value using internal data, and must handle varying acquisition costs for different user states, multi‑channel integration, and resource constraints.

Algorithm Flow

Offline, millions of ads are generated. The pipeline includes: (1) manual or algorithmic selection to keep the ad pool fresh; (2) personalized multi‑path recall to retrieve hundreds of candidate ads per user; (3) rule‑based filtering (e.g., deduplication, exposure frequency control); (4) value estimation via predictive models; (5) pricing algorithm that computes a bid, finally returning the ad and bid to the ADX.

This process borrows the multi‑path recall from recommendation systems but adds CTR bias handling and a pricing step after estimation.

Algorithm Implementation

Strict RT constraints (≤100 ms), multi‑channel diversity, and high traffic volume demand efficient models. Common ranking models include LR, GBDT, and DNN (DCN/WDL/MMOE). LR requires extensive feature engineering; GBDT trains quickly but may overfit categorical features; DNN handles diverse features via embeddings but incurs high memory and latency.

To unify models across channels, platform identifiers are used as features, and statistical metrics per platform are constructed, enabling a transition from single‑channel LR to a unified GBDT model, yielding noticeable business gains.

Problems and Reflections

Item‑based recall assumes watched videos reflect user preference, which is not always true, leading to mismatched recommendations and limited diversity for low‑frequency users. From a causal inference perspective, user silence stems from unsatisfactory recommendations, while retention arises from liked content. By building unbiased similarity metrics and matching low‑activity users to high‑activity “counterfactual” peers, we improve recommendation relevance.

Formal Problem Definition

Different scenarios impose distinct constraints on the pricing algorithm: daily bidding (budget, first‑activation cost, user‑value constraints) and burst (冲量) bidding (ample budget, first‑activation cost, user‑value constraints). Both are multi‑layer, multi‑channel optimization problems.

Single‑PID Solution for Burst Bidding

Under burst conditions, the bid is derived as CTR/CVR multiplied by a controllable coefficient α. A PID controller adjusts α to satisfy cost constraints while maximizing conversion.

Daily Bidding Solution

Daily bidding is modeled as a multi‑objective constrained maximization problem solved via Lagrangian duality, yielding a bid formula with parameters p and q. Two PID controllers regulate p and q, with an additional correction signal in the q controller to target final CPC, and a linear interpolation to capture coupling effects.

Pricing Engineering

The request flow: user request enters DSP management backend → logs (bidding, exposure, click) are parsed by Blink → error data stored in memory DB → offline platform generates delivery plan (cost, budget, parameters) → memory DB provides current pricing parameters → TPP platform performs ad recall and CTR estimation → final ad and bid are returned.

Engineering considerations include balancing theoretical support with machine and human resource constraints, leveraging internal and open‑source tools, and driving efficiency through A/B testing and layered optimization. Future directions involve extending CTR estimation to user‑value prediction, exploring reinforcement‑learning‑based pricing, and applying the framework to new user acquisition, hot‑show promotion, and membership scenarios.

Thank you for your attention.