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.
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.
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