GLINT-RU: Gated Lightweight Intelligent Recurrent Units for Sequential Recommender Systems

TL;DR

GLINT-RU addresses the computational burden of transformer-based sequential recommender systems by introducing a single-layer, gate-enhanced architecture that fuses a dense selective GRU with a linear attention expert through a parallel expert mixing block. The dense selective GRU captures local temporal patterns and fine-grained positional information, while the expert mixing block combines long-term dependencies and global interactions with linear complexity. The overall pipeline includes a dense selective GRU module, linear attention, and a gated MLP for output filtering, achieving $O((2k+12) N d^2)$ time complexity and high efficiency. Across three datasets (ML-1M, Amazon Beauty, Amazon Video Games), GLINT-RU yields higher Recall@10, MRR@10, and NDCG@10 and faster inference than baselines, demonstrating strong practical impact for SRS tasks.

Abstract

Transformer-based models have gained significant traction in sequential recommender systems (SRSs) for their ability to capture user-item interactions effectively. However, these models often suffer from high computational costs and slow inference. Meanwhile, existing efficient SRS approaches struggle to embed high-quality semantic and positional information into latent representations. To tackle these challenges, this paper introduces GLINT-RU, a lightweight and efficient SRS leveraging a single-layer dense selective Gated Recurrent Units (GRU) module to accelerate inference. By incorporating a dense selective gate, GLINT-RU adaptively captures temporal dependencies and fine-grained positional information, generating high-quality latent representations. Additionally, a parallel mixing block infuses fine-grained positional features into user-item interactions, enhancing both recommendation quality and efficiency. Extensive experiments on three datasets demonstrate that GLINT-RU achieves superior prediction accuracy and inference speed, outperforming baselines based on RNNs, Transformers, MLPs, and SSMs. These results establish GLINT-RU as a powerful and efficient solution for SRSs.

Figures (6)

• Figure 1: (a). Framework of proposed GLINT-RU. (b). Expert mixing block employs paralleled attention and GRU to effectively learn semantic features and fine-grained positional information. (c). Dense Selective GRU as the core part of the framework, deeply selects and aggregates the hidden states. (d). Gated MLP block is utilized to deeply filter the feed forward network.
• Figure 2: General process of expert mixing mechanism. GRU captures long-term dependencies with recurrent latent and fine-grained positional representations. Attention layer learns semantic features fro important item interactions.
• Figure 3: Impacts of kernel size $k$ of the temporal convolution on the performance of GLINT-RU and Mamba4Rec and the GPU Memory occupation of GLINT-RU.
• Figure 4: Impacts of kernel size $k$ of the temporal convolution on the training and inference time.
• Figure 5: Impacts of hidden size $d$ on the performance of GLINT-RU and state-of-the-art baselines.