Exploiting global and local item transition patterns for sequential recommendation

Bin Wu

doi:10.1016/j.datak.2026.102606

Exploiting global and local item transition patterns for sequential recommendation

2026-04-12

0

PRE

AI

DOI：10.1016/j.datak.2026.102606

Original

Original request for help

Save

Abstract

En 中文

Sequential recommendation aims to infer the next item that a user will visit, which captures dynamic user interests by modeling item transition patterns. While Transformer-based methods are a popular choice for this task, recent studies have shown that self-attention mechanisms’ time complexity is proportional to the square of the length of a user’s behavior sequence. Multilayer perceptron (MLP) based recommenders have shown promise in alleviating this efficiency issue, which extracts item transition patterns by some pure MLP blocks. Despite their effectiveness, three crucial limitations have not been well-addressed in previous methods: (1) they merely capture item transition patterns within each user action sequence, which fails to excavate transitional signals across different sequences; (2) these methods directly adopt the one-hot encoding technique to model long-term user preference, which is non-trivial to exploit collaborative signals; (3) they randomly select several uninteracted items as negatives, which are easy to discriminate and provide limited information for learning optimal parameters. In light of these limitations, we propose a novel Global-Local Sequential Recommender (GLSRec), which jointly models global and local item transition patterns for dynamic user interests. First, we construct a global item transition graph by reorganizing all users’ behavior sequences. Afterwards, we present a graph transition module upon this graph to distill global transitional signals into item embeddings, which can be viewed as a calibrator and guide the follow-up local MLP blocks. Furthermore, rather than selecting some portions of unobserved data as negatives, we design an interest-aware negative sample generator to dynamically synthesize hard negatives during training. Experimental results on four benchmark datasets demonstrate the superiority of GLSRec over recent state-of-the-art models. The source code of our implementation is publicly available at https://github.com/wubinzzu/GLSRec .

Keywords:

Sequential recommendation

Item transition patterns

Global-local modeling

Multilayer perceptron

Negative sampling

Exploiting global and local item transition patterns for sequential recommendation

Abstract

Journal

Organization