Research News

Sequential point-of-interest (POI) recommendations in niche cultural-tourism settings must capture users’ parallel interests and rapid intent shifts.

Therefore, an Osaka Metropolitan University research team conducted a study that proposes a transparent multi-task framework that combines a multi-slot user memory with a dynamic gating network to jointly predict the next POI and the next thematic intent. User histories are divided into interpretable slots, such as travel themes or railway lines, and an encoder state at each step is compared with these slots to identify the most relevant one. The gate’s weights indicate which preference slot drives each prediction, providing step-level explanations. 

Experiments on a real-world dataset of Japanese railway-stamp rallies demonstrate consistent improvements of 1.5–4.0 percentage points in Precision@5, NDCG@5, and theme-classification accuracy over strong baselines including GRU4Rec, SASRec, and single-slot memory models. Ablation analysis confirms that both the multi-slot memory and the gating mechanism contribute to performance gains. Visualization of gating patterns further reveals meaningful intent transitions, offering human-readable transparency valuable for cultural-tourism applications. The findings highlight the importance of interpretable memory and multi-task learning for accurate and explainable recommendations in specialized POI environments.

Paper information

Journal: IEEE Access
Title: Multi-Slot Memory With Dynamic Gating: A Multi-Task Framework for Interpretable Sequential Recommendation in Niche POI Scenarios
DOI: 10.1109/ACCESS.2025.3632668
Authors: Zhaoqi Ma, Jiansong Tang, Ryosuke Saga
Published:  13 November 2025
URL: https://doi.org/10.1109/ACCESS.2025.3632668

Contact

Ryosuke Saga
Graduate School of Informatics
Email: r.saga[at]omu.ac.jp

*Please change [at] to @.