Visual Perception–Driven Point-of-Interest Prediction via Multi-View Disentanglement and Spatial Encoding
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
The effective modeling of complex multidimensional interactions in spatial environments is a critical challenge in intelligent sensing and urban computing.Existing approaches often fail to simultaneously capture global structural dependencies and local spatial attributes when modeling spatial interaction intensity and location distribution patterns.Furthermore, representation entanglement during multi-source information fusion limits the capacity to characterize complex scenarios.To address these limitations, a graph-based representation learning framework that integrates spatial frequency information and positional encoding was proposed from the perspective of spatial perception modeling.Furthermore, a spatial representation model based on multi-view disentanglement and adaptive fusion was developed.Spatial interaction frequency was introduced to weight node relationships, thereby enhancing the expressiveness of graph structures in capturing significant interaction patterns.In parallel, positional encoding was incorporated into node representations to improve sensitivity to spatial distribution characteristics.Building upon this foundation, a multi-view disentangled representation learning framework was constructed, in which global interaction dependencies, temporal behavioral patterns, and spatial proximity features were modeled separately.A contrastive learning strategy was employed to enforce consistency across views while promoting complementary information extraction.Subsequently, an adaptive fusion mechanism was utilized to generate unified spatial representations.Experimental results on real-world spatial interaction datasets demonstrate that superior performance was achieved across multiple evaluation metrics compared to baseline methods, indicating that prediction accuracy for spatial targets is significantly improved.These findings suggest that the integration of graph representation learning with multi-view disentanglement enhances the model's capability to capture complex spatial scenarios and offers a promising approach for visual perception-driven spatial intelligence applications.
Fields of Study & Subjects
Advanced Graph Neural Networks
Human Mobility and Location-Based Analysis
Urban Design and Spatial Analysis
Publication details
- Journal / source
- Traitement du signal
- Volume
- 43
- Issue
- 1
- Publication date
- 28 February 2026
- Language
- EN