Mamba4Rec: Efficient Sequential Recommendation with Selective State Space Models (arXiv 2403.03900)

Forum topic · 小凯 · 2026-07-05

Summary

Mamba4Rec is a March 2024 arXiv paper (arXiv:2403.03900) by Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee that explores applying selective state space models (SSMs), popularized by the Mamba architecture, to sequential recommendation. The work aims to address the efficiency, scalability, and user-intent modeling challenges that Transformer-based sequential recommenders face due to quadratic attention complexity. By leveraging Mamba's selective state space mechanism, Mamba4Rec seeks linear-time sequence modeling with effective long-range dependency capture over user interaction histories. This forum post catalogs the paper's metadata, positions it within the sequential recommendation literature (alongside BERT4Rec, EAGER, and Transformer-based recommenders), and summarizes the typical pipeline of such systems: input representation, core modeling modules, learning strategies, and inference strategies. Quantitative results should be verified against the original PDF, as the post is based on abstract-level metadata.

Mamba4Rec: Towards Efficient Sequential Recommendation with Selective State Space Models

Paper: Mamba4Rec: Towards Efficient Sequential Recommendation with Selective State Space Models (arXiv, March 2024)
Authors: Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, James Caverlee
Category: Sequential Recommendation

Key points

Mamba4Rec applies selective state space models (SSMs) — the core mechanism behind the Mamba architecture — to sequential recommendation, replacing or complementing Transformer-based backbone models.
The motivation is that Transformer recommenders suffer from quadratic attention complexity with sequence length, which limits efficiency and scalability in large-scale recommendation systems with long user behavior histories.
Selective SSMs offer linear-time sequence modeling while retaining the ability to selectively propagate or forget information, making them a candidate for efficient modeling of long interaction sequences.
The work sits in the broader lineage of sequential recommenders such as BERT4Rec, multi-behavior Transformers, and generative recommenders, where the core tensions are sparse user behavior, enormous item catalogs, and multi-objective trade-offs.

Context within sequential recommendation

Sequential recommendation has evolved from matrix factorization to deep CTR models, sequence Transformers, and now instruction-following / generative recommenders. Each generation balances the efficiency–effectiveness–maintainability triangle:

Self-attention models capture rich dependencies but scale poorly with sequence length.
Mamba-style selective SSMs promise linear complexity with selective memory, addressing latency and compute budgets that matter for production systems.

Caveats

This post is based on abstract-level metadata; specific benchmark numbers, datasets, and ablation results should be verified against the original PDF.

Related entries

BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations
EAGER: Two-Stream Generative Recommender
Efficient On-Device Session-Based Recommendation (ACM TORS)
How to Index Item IDs for Recommendation Foundation Models (P5, SIGIR)
LLMCDSR: Cross-Domain Sequential Recommendation with LLMs
Multi-Behavior Sequential Transformer Recommender (SIGIR 2024)