RaSeRec: Retrieval-Augmented Sequential Recommendation

TL;DR

RaSeRec introduces a Retrieval-Augmented Sequential Recommendation framework that leverages a dynamic memory bank to address preference drift and long-tail recall. It combines a collaborative-based pre-training stage (learning to recommend and retrieve) with a retrieval-augmented fine-tuning stage (learning to use retrieved memories via a RAM with dual-channel cross-attention). The approach yields consistent improvements over state-of-the-art baselines across three datasets and multiple backbones, demonstrates robustness to noise, and offers insights into memory-bank partitioning and hyperparameter sensitivity. The work advances practical SeRec by enabling explicit memory retrieval to complement implicit model memory, with potential for broad applicability and future active-retrieval extensions.

Abstract

Although prevailing supervised and self-supervised learning augmented sequential recommendation (SeRec) models have achieved improved performance with powerful neural network architectures, we argue that they still suffer from two limitations: (1) Preference Drift, where models trained on past data can hardly accommodate evolving user preference; and (2) Implicit Memory, where head patterns dominate parametric learning, making it harder to recall long tails. In this work, we explore retrieval augmentation in SeRec, to address these limitations. Specifically, we propose a Retrieval-Augmented Sequential Recommendation framework, named RaSeRec, the main idea of which is to maintain a dynamic memory bank to accommodate preference drifts and retrieve relevant memories to augment user modeling explicitly. It consists of two stages: (i) collaborative-based pre-training, which learns to recommend and retrieve; (ii) retrieval-augmented fine-tuning, which learns to leverage retrieved memories. Extensive experiments on three datasets fully demonstrate the superiority and effectiveness of RaSeRec. The implementation code is available at https://github.com/HITsz-TMG/RaSeRec.

Figures (8)

• Figure 1: Comparison between Vanilla, SSL-Augmented, and Retrieval-Augmented SeRec paradigms, where their system working flows are illustrated from left to right, respectively. $s_u$ denotes the user sequence, while $s_u^\prime, s_u^{\prime\prime}$ represent two augmented views. SeqEnc denotes the sequence encoder, referring to §\ref{['sec:trans_sr']} for more technical details.
• Figure 2: The overall system framework of the proposed RaSeRec. The upper layer illustrates the workflow of collaborative-based pre-training while the bottom layer shows the workflow of retrieval-augmented fine-tuning.
• Figure 3: The architecture of Retrieval-Augmented Module, where $\tilde{\mathbf{h}}$ is the augmented user representation.
• Figure 4: Performance comparison over different item groups between RaSeRec and DuoRec. The bar represents NDCG@5, while the line represents the performance improvement percentage of RaSeRec compared with DuoRec.
• Figure 5: Performance comparison when ablating different partitions of the memory bank $\mathcal{M}$. The bar represents HR@5 or NDCG@5, while the line represents the percentage of performance degradation compared to the "All".