Smoothed Intent Learning for Sequential Recommendation

Intent-aware sequential recommendation (ISR) predicts next-item interactions by capturing the implicit intents behind users’ historical behavior sequences. Although existing ISR methods have achieved promising performance, we argue that two smoothness issues remain insufficiently addressed. First, existing methods often overemphasize short-term intents, making it difficult to smoothly balance users’ short- and long-term preferences. Second, hard positive–negative boundaries in intent contrastive learning may introduce misleading contrastive signals and disrupt the semantic continuity of user behaviors. To tackle these issues, we propose S moothed I ntent L earning for sequential Rec ommendation ( SILRec ), which improves the smoothness of short- and long-term intents modeling as well as intent contrastive samples. Specifically, we design a SILRec encoder with a multi-scale exponential moving average module and a frequency-domain module to model user intents across different temporal and spectral patterns. We further introduce boundary-smoothed contrastive learning based on representation interpolation, which constructs smoother transition regions between intent representations and alleviates the rigidity of hard contrastive sample boundaries. Comprehensive experiments on six conventional real-world datasets with thirteen competitors demonstrate the effectiveness of SILRec. We further conduct an additional evaluation on the KuaiRec short-video dataset to examine its applicability to more dynamic recommendation scenarios. Additional analyses from multiple perspectives, including controlled noise perturbation, long-tail performance, and intent visualization, provide further evidence for the proposed design. Our codes and datasets are available at https://github.com/syf1844803351/SILRec .