MOBIUS: Towards the Next Generation of Query-Ad Matching in Baidu's Sponsored Search

Miao Fan, Jiacheng Guo, Shuai Zhu, Shuo Miao, Mingming Sun, Ping Li

Abstract

Baidu’s MOBIUS project revolutionizes sponsored search by unifying the traditionally separate "relevance matching" and "business ranking" layers into a single, high-efficiency system. By employing a teacher-student active learning framework and state-of-the-art approximate nearest neighbor (ANN) search, MOBIUS predicts click-through rates (CTR) for billions of ads in milliseconds. This approach addresses the "blind spot" inherent in multi-stage architectures, significantly boosting both revenue and retrieval efficiency.

1. Introduction: The Challenge of Decoupled Architectures

In traditional sponsored search engines, such as Baidu’s "Phoenix Nest," ads are processed through a multi-stage funnel designed to manage computational complexity. The Matching Layer screens billions of candidates for semantic relevance to the user's query, while the Ranking Layer takes the top few thousand candidates and sorts them based on business metrics like Cost Per Mille (CPM) and Return on Investment (ROI).

The traditional multi-stage funnel architecture in sponsored search, where matching and ranking are decoupled.

This decoupled architecture creates a significant "blind spot." The matching layer is typically agnostic to business value, and the ranking layer can only optimize what the matching layer permits. Consequently, high-value ads may be filtered out early if they lack strong keyword overlap, even if they are highly likely to be clicked. Furthermore, traditional models often suffer from "click history bias," overestimating the value of popular ads regardless of query relevance.

2. The MOBIUS Framework: Unifying Matching and Ranking

MOBIUS collapses this traditional funnel by teaching the matching layer to predict CTR directly, effectively bringing business logic to the retrieval frontline.

The MOBIUS unified architecture, which integrates multi-objective optimization directly into the retrieval of billions of ad candidates.

2.1 Active-Learned CTR Model

To train a model on billions of query-ad pairs, MOBIUS utilizes a teacher-student active learning framework. Since historical click data is sparse for "tail" queries, the system generates synthetic query-ad pairs via cross-join operations. A high-precision relevance model (the "Teacher") grades these synthetic pairs. A large-scale Deep Neural Network (the "Student") is then trained on "bad cases"—pairs with high predicted CTR but low relevance as flagged by the Teacher. This ensures the model learns to avoid irrelevant but popular ads.

2.2 High-Efficiency Retrieval

Predicting CTR for billions of ads in real-time requires significant optimization. MOBIUS employs a two-tower DNN architecture to generate 96-dimensional embeddings for queries and ads. The ranking problem is formalized as a Maximum Inner Product Search (MIPS). To manage memory and latency, Optimized Product Quantization (OPQ) compresses these vectors, allowing the system to search billions of ads with only 5% of the original memory footprint.

3. Experimental Results

The system was evaluated using Baidu’s internal search logs, comprising 80 billion training samples and 10 billion test records. Deployment of MOBIUS-V1 yielded the following results:

  • Revenue Growth: CPM increased by 3.8% on the Baidu mobile app and 3.5% on PC.
  • Ad Coverage: The ad coverage rate increased from 7.3% to 40.5%, demonstrating a superior ability to find relevant, high-value ads.
  • System Efficiency: Query latency for high-value ads was reduced by 75%, from 120ms to 30ms.
  • Relevance: Human expert evaluations confirmed that the active learning approach significantly improved semantic relevance compared to baseline models.

4. Conclusion and Future Work

MOBIUS demonstrates that unifying retrieval and ranking objectives is feasible at scale. By treating the retrieval stage as an estimator of business value, Baidu has achieved significant gains in both user experience and monetization. Future work will focus on incorporating additional business targets, such as ROI, directly into the matching layer and exploring GPU-based hardware acceleration for ANN search.

Summary updated

Abstract

Baidu runs the largest commercial web search engine in China, serving hundreds of millions of online users every day in response to a great variety of queries. In order to build a high-efficiency sponsored search engine, we used to adopt a three-layer funnel-shaped structure to screen and sort hundreds of ads from billions of ad candidates subject to the requirement of low response latency and the restraints of computing resources. Given a user query, the top matching layer is responsible for providing semantically relevant ad candidates to the next layer, while the ranking layer at the bottom concerns more about business indicators (e.g., CPM, ROI, etc.) of those ads. The clear separation between the matching and ranking objectives results in a lower commercial return. The Mobius project has been established to address this serious issue. It is our first attempt to train the matching layer to consider CPM as an additional optimization objective besides the query-ad relevance, via directly predicting CTR (click-through rate) from billions of query-ad pairs. Specifically, this paper will elaborate on how we adopt active learning to overcome the insufficiency of click history at the matching layer when training our neural click networks offline, and how we use the SOTA ANN search technique for retrieving ads more efficiently (Here ``ANN'' stands for approximate nearest neighbor search). We contribute the solutions to Mobius-V1 as the first version of our next generation query-ad matching system.
BibTeX
@inproceedings{fan2019mobius,
  author = {Miao Fan and Jiacheng Guo and Shuai Zhu and Shuo Miao and Mingming Sun and Ping Li},
  title = {{MOBIUS: Towards the Next Generation of Query-Ad Matching in Baidu's Sponsored Search}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2019},
  doi = {10.1145/3292500.3330651},
}
Figure