@article{Yang2026, 
author = {Xingyao Yang and Yanfu Wu and Zulian Zhang and Jiong Yu and Zhiqiang Zhong and Yu Chen},
title = {Fine-Grained Long and Short-Term Preference Sequential Recommendation with Contrastive Learning},
year = {2026},
journal = {Journal of Xinjiang University(Natural Science Edition in Chinese and English)},
volume = {43},
number = {2},
pages = {156-168},
keywords = {recommender system, sequential recommendation, contrastive learning, self-attention mechanism, gated recurrent unit},
url = {https://www.sciopen.com/article/10.13568/j.cnki.651094.651316.2025.01.18.0001},
doi = {10.13568/j.cnki.651094.651316.2025.01.18.0001},
abstract = {Sequence recommendation aims at item recommendation using users' long and short-term preferences, but most sequence recommendation systems face problems such as insufficient learning power and inadequate fusion of long and short-term preferences. Aiming at the above problems, this paper proposes a fine-grained long and short-term preference sequence recommendation method based on contrastive learning. 1) To address the problem of insufficient long and short-term preference fusion, this paper proposes a long and short-term preference learning layer and a long and short-term preference fusion layer. Firstly, it splits the user behaviour sequence into multi-period sessions and extracts the user's short-term preference in each session by using gated recurrent units, and then fuses the short-term preference sequences to capture the user's long-term preference through the multi-head attention mechanism. Finally, the long-term and short-term preferences are fused adaptively based on the time span to obtain a more representative and comprehensive preference representation. 2) Aiming at the problem of insufficient learning power due to data sparsity, a preference representation comparison learning task is designed to introduce agent user preferences for comparison learning to achieve more accurate preference recommendation. The experimental results show that: compared to the sub-optimal methods, the model improves the Hit@20 metric by 9.84%, 6.40%, and 1.52%, and the MAP@20 metric by 22.64%, 2.42%, and 6.42% on three public datasets, respectively, demonstrating the effectiveness of the proposed method.}
}