OptMSM: Optimizing Multi-Scenario Modeling for Click-Through Rate Prediction

Xing Tang, Yang Qiao, Yuwen Fu, Fuyuan Lyu, Dugang Liu, Xiuqiang He

Abstract

A large-scale industrial recommendation platform typically consists of multiple associated scenarios, requiring a unified click-through rate (CTR) prediction model to serve them simultaneously. Existing approaches for multi-scenario CTR prediction generally consist of two main modules: a scenario-aware learning module that learns representations from input features, and a scenario-specific prediction module. However, most approaches neglect the latter, resulting in increased parameter size and performance bottlenecks. To address these issues, we propose OptMSM (Optimizing Multi-Scenario Modeling). First, we introduce a scenario-enhanced learning module that partitions input features and imposes an orthogonality constraint on shared embeddings to facilitate disentanglement. Second, we introduce a scenario-specific hypernetwork in the prediction module to capture interactions within each scenario more effectively. Extensive offline experiments and online A/B tests demonstrate the effectiveness of OptMSM.

Introduction

In modern industrial recommendation systems, users interact with various scenarios within a single application—such as a general homepage, specific product pages, or investment portfolios. While training individual models for each scenario is computationally expensive, unified Multi-Scenario Modeling (MSM) often suffers from "negative transfer," where data from one scenario degrades the performance of another. OptMSM is designed to mitigate these challenges by optimizing both the representation learning and the final prediction stages.

The OptMSM Framework

OptMSM optimizes the multi-scenario pipeline through three primary components: feature partitioning, scenario-enhanced learning, and adaptive prediction.

1. Feature Partitioning and Enhanced Learning

Unlike standard models that treat all features uniformly, OptMSM explicitly splits input features into scenario-specific (e.g., page position) and scenario-shared (e.g., user demographics).

To prevent interference between scenarios, the framework employs an orthogonality constraint. This mathematical constraint forces the internal representations of different scenarios to be distinct in vector space. By "untangling" these representations, the model can effectively separate common knowledge from scenario-specific nuances, directly addressing the negative transfer problem.

The overall architecture of the OptMSM framework, highlighting the feature partition module and the scenario-enhanced learning module with orthogonality constraints.

2. Scenario-Specific Prediction via Hypernetworks

Most existing research focuses solely on learning better feature embeddings. OptMSM goes further by optimizing the prediction architecture itself. It utilizes a Hypernetwork—a small, auxiliary network that dynamically generates the weights for the main prediction network based on the current scenario. This allows the model to adapt its decision-making logic instantly depending on the user's current context within the app, providing a more efficient form of personalization than simply adding more static layers.

The scenario-specific prediction module utilizing a HyperNetwork to dynamically generate model parameters for each scenario.

Experimental Results

Offline Benchmarks

The researchers validated OptMSM on two public datasets (Ali-CCP and AliExpress) and a massive production dataset from Tencent Licaitong. OptMSM consistently outperformed state-of-the-art baselines like MMoE, PLE, and STAR in terms of Area Under the Curve (AUC) across nearly all tested scenarios.

Online Impact

In a real-world deployment on the Tencent Licaitong financial platform, OptMSM delivered significant business impact:

  • CTR Lift: An increase of +1.42% to +1.76% in click-through rates.
  • Revenue Impact: A substantial +5.82% to +7.10% lift in Purchase Amount Per Mille (PAPM), proving that architectural improvements in CTR modeling translate directly to financial performance.
  • Operational Efficiency: Despite its advanced architecture, the training overhead increased by only ~4.5%, making it highly viable for large-scale production environments.

Conclusion

OptMSM represents a significant step forward in multi-scenario modeling by balancing shared learning with scenario-specific adaptation. By combining feature disentanglement through orthogonality constraints with dynamic parameter generation via hypernetworks, it provides a robust and efficient framework for complex, multi-environment recommendation systems.

Summary updated

Abstract

A large-scale industrial recommendation platform typically consists of multiple associated scenarios, requiring a unified click-through rate (CTR) prediction model to serve them simultaneously. Existing approaches for multi-scenario CTR prediction generally consist of two main modules: i) a scenario-aware learning module that learns a set of multi-functional representations with scenario-shared and scenario-specific information from input features, and ii) a scenario-specific prediction module that serves each scenario based on these representations. However, most of these approaches primarily focus on improving the former module and neglect the latter module. This can result in challenges such as increased model parameter size, training difficulty, and performance bottlenecks for each scenario. To address these issues, we propose a novel framework called OptMSM (\textbf{Opt}imizing \textbf{M}ulti-\textbf{S}cenario \textbf{M}odeling). First, we introduce a simplified yet effective scenario-enhanced learning module to alleviate the aforementioned challenges. Specifically, we partition the input features into scenario-specific and scenario-shared features, which are mapped to specific information embedding encodings and a set of shared information embeddings, respectively. By imposing an orthogonality constraint on the shared information embeddings to facilitate the disentanglement of shared information corresponding to each scenario, we combine them with the specific information embeddings to obtain multi-functional representations. Second, we introduce a scenario-specific hypernetwork in the scenario-specific prediction module to capture interactions within each scenario more effectively, thereby alleviating the performance bottlenecks. Finally, we conduct extensive offline experiments and an online A/B test to demonstrate the effectiveness of OptMSM.
BibTeX
@inproceedings{tang2023optmsm,
  author = {Xing Tang and Yang Qiao and Yuwen Fu and Fuyuan Lyu and Dugang Liu and Xiuqiang He},
  title = {{OptMSM: Optimizing Multi-Scenario Modeling for Click-Through Rate Prediction}},
  booktitle = {ECML/PKDD},
  year = {2023},
  doi = {10.1007/978-3-031-43427-3\_34},
}
Figure