UniFormer: Efficient and Unified Model-Centric Scaling for Industrial Recommendation

Industrial recommender systems have traditionally relied on component-centric scaling, where individual modules (behavior, feature interaction, task modeling) are scaled independently. This fragmented approach often leads to sub-optimal resource utilization and limits the model's ability to capture complex, cross-component dependencies. The authors seek to transition toward a unified model-centric scaling paradigm that can efficiently scale across the entire modeling space while maintaining the low-latency requirements of large-scale production environments.

UniFormer introduces a two-part architecture: the Feature-space Interaction Module (FIM) and the Task-space Interaction Module (TIM).

• Tokenization: The model uses a semantic-based tokenization scheme that groups features by their dependency on the target item. This allows for user-item decoupling, where request-level features are computed once and reused across multiple candidates, significantly accelerating inference.
• Unified Interaction: FIM handles interactions between sequential and non-sequential features using multi-sequence cross-attention to prevent preference collapse, while TIM integrates these high-order feature representations with task-specific signals.
• Scalability: The framework employs dedicated multi-view Feed-Forward Networks (FFNs) for different components, allowing for flexible parameter allocation that prevents any single module from becoming a bottleneck.

UniFormer was evaluated through extensive offline experiments and online A/B testing in Kuaishou’s production environments. The results demonstrate that the unified scaling approach consistently improves user engagement metrics. Specifically, in Kuaishou and Kuaishou Lite scenarios, the model achieved significant gains in both App Stay Time (+0.101% to +0.260%) and Watch Time (+0.729% to +1.113%). These improvements validate the effectiveness of co-scaling feature and task spaces within a single, unified architecture.

Why It Matters

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.