Multi-Scenario Recommendation Model

This paper presents a scenario-adaptive and self-supervised model (SASS) for multi-scenario personalized recommendation, addressing challenges in content recommendation across diverse scenarios. The model integrates global sample expansion through contrastive learning and scene-specific adaptation via multi-layer gate mechanisms. Key contributions include a two-stage training framework, scene-aware parameter transfer, and cross-domain modeling between content and commerce domains. Experimental results demonstrate improved recommendation performance across five core scenarios in Taobao's 'SuoSuo' platform.

The framework combines scene-contrastive pre-training with adaptive transfer modules to handle data sparsity and domain differences. It enables one-model deployment across multiple scenarios while maintaining scene-specific personalization. Applications extend to cold start mitigation, cross-domain modeling, and interactive engagement optimization.