Existing sequential recommendation systems face challenges in simultaneously capturing semantic patterns and temporal dynamics, particularly in cross-domain and data-scarce scenarios. We present TDSR, a dual-domain framework that combines semantic representation learning with temporal pattern modeling. The framework integrates three key components: 1) LSH-enhanced semantic mapping with collaborative denoising for robust embeddings, 2) adaptive temporal modeling using time-aware attention mechanisms, and 3) cross-domain fusion with orthogonal constraints and adversarial alignment to preserve domain characteristics. Experimental results demonstrate TDSR’s superior performance over state-of-the-art methods, especially under data sparsity, revealing new insights into semantic-temporal interaction for recommendation systems.