北京信息科技大学学报（自然科学版）2025，Vol.40Issue(6)：90-98,9.DOI:10.16508/j.cnki.11-5866/n.2025.06.010
结合周期感知去噪与频率时变建模的序列推荐模型
A sequential recommendation model with cycle-aware denoising and frequency-temporal modeling
摘要
Abstract
To address the challenges of insufficient periodic pattern mining,noise interference,and difficulty in capturing dynamic patterns in user behavior sequence prediction,a multi-module collaborative prediction framework was proposed.Firstly,a cycle-aware denoising module that includes sub-modules for commodity periodicity perception and noise-like filtering was established,and periodic feature embeddings were generated by mining periodic patterns in user behaviors.Periodic redundant noise-like data was suppressed more precisely by integrating temporal interval features,thereby addressing the problem that traditional denoising methods tend to mistakenly remove valid interest signals and providing clean behavioral sequence inputs for subsequent modeling.Secondly,the frequency-temporal modeling module was employed to convert time-domain sequences into frequency-domain representations using the fast Fourier transform.The local dominant frequency features were extracted through sliding windows and mapped into enhanced vectors,overcoming the limitation of time-domain modeling in capturing dynamic frequency changes,thus enabling fine-grained characterization of the evolutionary patterns of user interests.These two core modules form a progressive and collaborative"noise purification-dynamic enhancement"link,collectively contributing to the improvement of user interest prediction accuracy.Finally,the Transformer module based on a sparse attention mechanism was used to predict the enhanced sequence.Experiments on the Rec-Tmall and UserBehavior public datasets demonstrate that the proposed method significantly outperforms multiple baseline models in core metrics such as NDCG@10,AUC,HIT@5,and MRR,effectively improving the accuracy and robustness of user interest prediction.关键词
序列推荐/周期模式挖掘/噪声过滤/快速傅里叶变换/稀疏注意力Key words
sequential recommendation/periodic pattern mining/noise filtering/fast Fourier transform(FFT)/sparse attention分类
信息技术与安全科学引用本文复制引用
李光耀,吕学强..结合周期感知去噪与频率时变建模的序列推荐模型[J].北京信息科技大学学报（自然科学版）,2025,40(6):90-98,9.基金项目
国家自然科学基金项目(62202061,62171043) （62202061,62171043）
北京市自然科学基金项目(4232025) （4232025）
北京市教委科研计划科技一般项目(KM202311232002) （KM202311232002）
