ADR-Rec: Adaptive disentanglement for cross-domain sequential recommendation with cross attention gating mechanisms

Cross-domain sequential recommendation (CDSR) aims to predict users’ next-item interactions by utilizing behavioral sequences across multiple domains. Traditional approaches often model user preferences separately in each domain and then transfer the learned knowledge, which fails to fully capture the rich information embedded in cross-domain item sequences (cross-sequences). Recent works attempt to address this limitation, but challenges remain. First, due to the overlap between cross-sequences and individual sequences, it is essential to disentangle domain-specific information from shared information. Second, the importance of cross-sequence information varies over time, requiring dynamic adjustment of its contribution. To tackle these issues, we propose a novel framework that separates user preferences into domain-specific and cross-domain representations, and dynamically adjusts their contributions at each time step. A domain discriminator with a gradient reversal layer extracts cross-domain preferences by learning domain-invariant representations, while the domain-specific preferences retain exclusive information. A gating layer is then used to modulate the influence of each preference type over time. Experiments on two real-world CDSR datasets show that our method significantly outperforms state-of-the-art baselines, validating the effectiveness of the proposed disentanglement and dynamic gating mechanisms.