Abstract:
Cross-domain sequential recommendation alleviates data sparsity by transferring source domain information from a source domain, but existing methods remain limited in fine-grained temporal modeling and cross-domain feature alignment. To address the above issues, this paper proposes a cross-domain sequential recommendation model with time-aware and orthogonal mapping, namely CDSR-TAOM. With orthogonal mapping constraints, the model integrates latent user features transferred from the source domain with fine-grained temporal information in the target domain, thereby improving the overall performance of cross-domain sequential recommendation. The CDSR-TAOM model consists of two core modules. 1) The time-interval-aware module introduces an improved attention mechanism to capture both the sequential dependencies and time interval features of user interaction sequences. 2) The orthogonal cross-domain mapping module is built on the modeling assumption of shared overlapping features and domain-specific features. It adopts orthogonal spatial transformation to decouple domain features, preserves the norm and inner-product structure of features during transfer, and effectively alleviates the problem of cross-domain feature confusion. Experiments are conducted on three cross-domain recommendation tasks constructed from the Books, CDs, Movies subsets of the public Amazon review dataset. Experimental results show that the proposed CDSR-TAOM model achieves superior
HR@
N and
NDCG@
N performance compared with baseline models across all three tasks. Further ablation experiments verify the effectiveness of the two core modules. The model integrates dynamic temporal modeling and cross-domain feature structure preservation, providing a feasible and effective solution for cross-domain sequential recommendation under data sparsity scenarios.