M $^{2}$ Rec: M ulti-scale M amba for Efficient Sequential Rec ommendation

Sequential recommendation tasks, aimed at predicting future user preferences based on their historical behavior, present substantial challenges due to the dynamic nature of user preferences and the intricacies of modeling temporal patterns. While Transformer-based methodologies in prior research struggled with efficiency concerns, Mamba-based strategies have emerged as a viable alternative. However, these approaches grapple with critical hurdles, including the complexities of capturing multi-scale periodic patterns (e.g., weekly, monthly behaviors), mitigating the adverse effects of data noise on prediction accuracy, and effectively modeling user preference contexts while aligning diverse modal sources. In response to these challenges, we introduce an efficient Mamba-based framework that leverages the power of Fast Fourier Transforms (FFTs) and the contextual insights offered by Large Language Models (LLMs), known asM$^{2}$Rec. Furthermore, we incorporate a gate mechanism to integrate a variety of modal information. Specifically, we employ an adaptive FFT to accurately capture multi-scale periodic user behavior while diminishing the impact of noise. Moreover, a novel gate mechanism aligns LLM embeddings with an enhanced Mamba encoder, thereby improving the model's understanding of user preferences through contextual semantics provided by LLMs. Through extensive evaluation on four real-world datasets, our proposed approach (M$^{2}$Rec) demonstrates superior performance across three pivotal metrics, surpassing all current state-of-the-art benchmarks. Sequential recommendation systems aim to predict users’ next preferences based on their interaction histories, but existing approaches face critical limitations in efficiency and multi-scale pattern recognition. While Transformer-based methods struggle with quadratic computational complexity, recent Mamba-based models improve efficiency but fail to capture periodic user behaviors, leverage rich semantic information, or effectively fuse multimodal features. To address these challenges, we proposeM$^{2}$Rec, a novel sequential recommendation framework that integrates multi-scale Mamba with Fourier analysis, Large Language Models (LLMs), and adaptive gating. First, we enhance Mamba with Fast Fourier Transform (FFT) to explicitly model periodic patterns in the frequency domain, separating meaningful trends from noise. Second, we incorporate LLM-based text embeddings to enrich sparse interaction data with semantic context from item descriptions. Finally, we introduce a learnable gate mechanism to dynamically balance temporal (Mamba), frequency (FFT), and semantic (LLM) features, ensuring harmonious multimodal fusion. Extensive experiments demonstrate thatM$^{2}$Recachieves state-of-the-art performance, improving Hit Rate@10 by 3.2% over existing Mamba-based models while maintaining 20% faster inference than Transformer baselines. Our results highlight the effectiveness of combining frequency analysis, semantic understanding, and adaptive fusion for sequential recommendation. Code and datasets are available at:https://github.com/CoderPowerBeyond/M2Rec.