Purify and Generalize: Efficient Dual-End Adapters for Sequential Recommendation

Abstract

Despite rapid advances in backbone architectures for sequential recommendation, existing approaches still suffer from two challenges. First, user behavior sequences contain noisy interactions that corrupt input representations and mislead the modeling process. Second, models often converge to sharp minima that overfit to popularity patterns, leading to unstable predictions and poor generalization. Although existing methods explore various techniques to mitigate these challenges, they often require model-specific designs or tailored modifications and rely on uniform transformations that lack adaptive control or instance-aware mechanisms. To address these issues, we present a novel Dual-End Adapter for Sequential Recommendation (DEASRec), which employs lightweight adapters as flexible plugins without modifying backbone architectures. Specifically, we first design a Polar Vector Modulation (PVM) at the input end, projecting embeddings into polar complex space and applying subspace-adaptive bounded modulation that decouples magnitude and phase for independent noise filtering. Then we propose Adaptive Stochastic Regularization (ASR) at the output end, injecting instance-aware Gaussian perturbations during training to encourage convergence toward flat minima. PVM and ASR form a theoretically grounded purify-then-generalize pipeline that addresses input noise and output overfitting. Extensive experiments on four real-world datasets demonstrate that DEASRec can be seamlessly integrated into different backbone architectures to achieve stable performance improvements with low computational cost. DEASRec also achieves competitive performance against state-of-the-art baselines.