Sequential Recommendation Model Based on Mamba2 and Adaptive Time-frequency Analysis
HU Chunyang
LIU Huanhuan
GE Bin
WEI Zhongliang
HUO Zhenge
Abstract：To address the limitations of self-attention-based Transformer models in sequential recommendation tasks,including insufficient dynamic interest capture and quadratic computational complexity growth with sequence length,a sequential recommendation model based on Mamba2 and adaptive time-frequency analysis(M2ATFSRec)was proposed.The model was designed to enhance dynamic interest modeling capability while reducing computational complexity and improving recommendation accuracy.Firstly,adaptive time-frequency analysis was employed to extract time-frequency features from user historical behavior sequences,explicitly encoding multi-scale periodic patterns of interests.Secondly,Mamba2's selective state space mechanism was utilized to achieve efficient dynamic interest modeling for long sequences.The M2ATFSRec was experimentally evaluated on three datasets,namely,the movielens 1 million ratings(MovieLens-1M),the Amazon beauty products(Amazon-Beauty),and the Amazon video games(Amazon-Video-Games).In terms of the normalized discounted cumulative gain(NDCG)metric,it was found that M2ATFSRec achieved improvements of 6.42%,22.76%,and 33.22%respectively compared to towards efficient sequential recommendation with selective state space model(Mamba4Rec).The model had better recommendation performance in long sequence scenarios.
Keywords：recommendation systemssequential recommendationMamba2data sparsityfrequency-domain fusionlong-short term preferences
Publication Date：2025-09-20
Online Publishing Date：2025-09-24(First online date of this platform, not the publication date of the document)
Pages：6( 405-410 )