Experimental evaluation demonstrates enhanced prediction accuracy in cross-domain sequential recommendation, indicating improved modeling of rapid user interest drift.
Key Points
To address static time-decay assumptions and disruptive negative sampling in cross-domain sequential recommendation by dynamically modeling user interest drift and maintaining cross-domain structural consistency.
Designed DiffCDSR, integrating a Position Re-weighting module that dynamically recalculates position importance to track short-term interest shifts.
Implemented a diffusion-guided mechanism to generate structure-preserving, semantically perturbed negative samples for contrastive learning.
Evaluated the framework on three standard cross-domain sequential recommendation benchmarks using NDCG@10 and HR@10 metrics.
DiffCDSR outperformed existing state-of-the-art baseline models across benchmark datasets in both NDCG@10 and HR@10 metrics.
Ablation experiments verified that both the position re-weighting module and the diffusion-guided negative sampling independently improved recommendation performance.