STMamba-GC: Spatiotemporal Mamba with graph contrastive learning for next POI recommendation.

Area of Science:

• Location-based services
• Artificial Intelligence
• Data Mining

Background:

• Next Point-of-Interest (POI) recommendation is crucial for personalized services.
• Existing methods struggle with sparse user data and dynamic preferences.
• Static spatiotemporal fusion limits adaptability in mobility contexts.

Purpose of the Study:

• To propose a novel network, Spatiotemporal Mamba with Graph Contrastive learning (STMamba-GC), for improved POI recommendation.
• To address challenges of sparse data, dynamic preferences, and static spatiotemporal fusion.
• To enhance user representation learning and adaptability to mobility contexts.

Main Methods:

• Constructing three complementary graph views: user transitions, semantic patterns, and geographic proximity.
• Employing graph augmentation with contrastive learning for robust preference extraction.
• Designing a spatiotemporal gated fusion Mamba module for dynamic spatiotemporal dependency balancing.
• Utilizing a multi-task learning mechanism for joint optimization of POI prediction and contrastive objectives.

Main Results:

• STMamba-GC demonstrated substantial improvements over state-of-the-art baselines on NYC and TKY datasets.
• Achieved relative gains of 10.065% in top-10 and 8.663% in top-20 accuracy on the NYC dataset.
• Obtained a 9.62% improvement in Mean Reciprocal Rank (MRR) on the TKY dataset.

Conclusions:

• The proposed STMamba-GC effectively addresses key challenges in Next POI recommendation.
• The dynamic spatiotemporal fusion and graph contrastive learning contribute to superior performance.
• STMamba-GC offers a promising advancement for location-based services and intelligent business decisions.