MOBIUS: A Next‑Generation Multi‑Objective Recall System for Baidu Sponsored Search

The article presents Baidu's next‑generation multi‑objective recall system, named "MOBIUS," which enhances the classic two‑stage recall‑ranking pipeline by incorporating business objectives like CPM directly into the recall layer, thereby improving overall system efficiency.

Innovations

Introduces CPM and other business metrics as recall criteria while maintaining relevance.

Integrates the traditional CTR prediction model into the recall stage, forming a novel commercial recall architecture.

System Architecture

The MOBIUS architecture consists of two core modules: a data‑augmentation module and a model‑training module. The data‑augmentation module generates training samples that highlight low‑relevance, high‑CTR (badcase) instances using an active‑learning teacher‑student framework.

Data Augmentation Process

Load a batch of click logs.

Construct query and ad sets from the batch.

Form all possible query‑ad pairs (N × M samples).

Score each pair with a relevance model and filter low‑relevance pairs.

Predict PCTR for the filtered pairs using a CTR model (T‑2).

Sample pairs based on PCTR, labeling them as badcase.

Combine the augmented badcase samples with original CTR training data, extending the binary classification to three classes (click, unclick, badcase).

Model Training Module

The model follows a classic dual‑tower design: user queries and ads are encoded into 96‑dimensional embeddings, split into three 32‑dimensional vectors, whose inner products produce three scores that are fed into a softmax layer to predict click, unclick, or badcase.

Online Retrieval

During online serving, query embeddings are used to retrieve high‑quality candidate ads via Approximate Nearest Neighbor (ANN) search (e.g., FAISS, HNSW). The system also supports Maximum Inner Product Search (MIPS) to embed business metrics directly into similarity calculations, and employs vector compression to reduce memory and storage costs.

Experimental Results

Offline experiments (see original paper) demonstrate significant improvements, and online deployment on Baidu PC and mobile platforms shows a substantial CPM increase within one week of monitoring.

Conclusion

By jointly optimizing relevance and monetization at the recall stage, MOBIUS achieves higher commercial value without sacrificing user experience, offering a promising direction for future ad retrieval systems.

Original source: https://zhuanlan.zhihu.com/p/146210155