Abstract: Aiming at the degradation of short sequence representations and semantic drift caused by data sparsity and random noise in sequential recommendation, a Dual-Layer Augmented robust sequential Recommendation model with dual-mixup decoupling regularization (DLARec) was proposed. Firstly, a structural repetition augmentation strategy was designed in the input enhancement stage. Recent real interaction segments were tiled in the idle padding positions of the sequence, and block identifiers combined with relative position encodings were adopted to clarify augmentation boundaries, to improve context signal density and alleviate representation degradation. Secondly, a dual-mixup decoupling module was constructed in the latent space regularization stage. Sequence features were interpolated with their randomly temporally-shuffled replicas to break the model’s overfitting to artificial augmentation layouts, and latent representations of different users within a mini-batch were linearly mixed to smooth decision boundaries, achieving implicit denoising and suppressing semantic drift. The model was evaluated on a total of six public benchmark datasets, including Beauty, Tool, and Music. The experimental results show that DLARec achieves state-of-the-art performance on Recall and Normalized Discounted Cumulative Gain (NDCG) metrics. Specifically, DLARec’s NDCG@10 metric outperforms SIGMA (SelectIve Gated MAmba) and TADA (Tail-Aware Data Augmentation) by 8.86% and 14.43% respectively on the extremely sparse Tool dataset and medium-long Music dataset. Through the collaborative mechanism of input-side structural augmentation and latent space feature smoothing, the model mitigates the interference caused by data sparsity and random noise, and improves the accuracy and robustness of recommendations.

Key words: Keywords: sequential recommendation, data augmentation, dual-mixup decoupling regularization, manifold mixup, robust recommendation

摘要： 针对序列推荐中因数据稀疏导致的短序列表征退化及随机噪声引发的语义漂移问题，提出一种双层增强的双重混合解耦正则化鲁棒序列推荐模型（DLARec)。首先，在输入增强阶段设计结构化重复增强策略，利用序列闲置填充位平铺近期的真实交互片段，并结合块标识与相对位置编码明确增强边界，以提升上下文信号密度并缓解表征退化；其次，在隐空间正则化阶段构建双重混合解耦模块，通过将序列特征与随机时序置换的副本进行插值，打破模型对人工增强布局的过拟合，并对批次内不同用户的隐层表征进行线性混合以平滑决策边界，实现隐式去噪并抑制语义漂移。在Beauty、Tool、Music等6个公开基准数据集上评估模型。实验结果表明，DLARec在召回率（Recall)和归一化折损累计增益（NDCG)指标上均取得最优表现。特别是在极度稀疏的Tool和中长序列的Music数据集上，DLARec模型的NDCG@10指标相较于SIGMA（SelectIve Gated MAmba)和TADA（Tail-Aware Data Augmentation)模型分别提升了8.86%和14.43%。该模型通过输入端结构化增强与隐空间特征平滑的协同机制，缓解数据稀疏与随机噪声带来的干扰，提升了推荐的准确性与鲁棒性。

关键词: 关键词: 序列推荐, 数据增强, 双重混合解耦正则化, 流形混合, 鲁棒推荐