Open Access

ARTICLE

Multi-Level Graph Signal Preservation for Sequential Recommendation with Selective State Spaces

Yitao Yang^1,2, Peng Wu^1,2,*, Xiaoming Zhang^3,*, Renjie Xu³, Yong Zhang³

1 School of Information Science and Engineering, Zhejiang Sci-Tech University, Hangzhou, China
2 Zhejiang Key Laboratory of Digital Fashion and Data Governance, Zhejiang Sci-Tech University, Hangzhou, China
3 National Key Laboratory of Intelligent Parallel Technology, Arms of the Army University, Beijing, China

* Corresponding Authors: Peng Wu. Email: ; Xiaoming Zhang. Email:

Computers, Materials & Continua 2026, 89(1), 92 https://doi.org/10.32604/cmc.2026.085505

Received 12 May 2026; Accepted 13 July 2026; Issue published 13 August 2026

Abstract

Existing graph-enhanced sequential recommendation methods typically adopt a unidirectional information flow, in which graph embeddings are injected into the sequential encoder only at the input stage, after which the graph signal is progressively diluted through multiple layers of deep processing. In this paper, the graph signal dilution phenomenon is analyzed systematically across three levels—the input, representation, and prediction layers—and the GSPRec model is proposed to address this issue. The core of GSPRec is the Graph-Sequence Collaborative Injection (GSCI) module, comprising three lightweight components: the Graph Confidence Gate (GCG) controls GCN smoothing via dimension-wise bounded interpolation; the Graph Residual Aggregation (GRA) restores diluted signals through a graph skip connection; and the Graph Collaborative Prediction (GCP) injects collaborative signals into prediction scores via a de-meaned shortcut. GSCI introduces only 194 learnable parameters in total and is equipped with a zero-damage initialization guarantee. The sequential encoder is further enhanced with independent dual Mamba instances and an adaptive path router. Experiments on four benchmark datasets—Food, Movie, Book, and Douban—demonstrate that GSPRec outperforms eight baselines across all 16 evaluation metrics, with relative improvements of 0.61%–3.16% over the strongest baseline and less than 4% additional training time. A layer-wise probing analysis directly confirms that the graph signal is diluted within the sequential encoder and that the GSCI components counteract this loss, while the learned injection strengths adapt across datasets.

---

Keywords

---