Subjective perception of urban streets enables subjective quantitative evaluation of street environments. Aiming at the spatio-temporal dynamics of subjective street perception and the variation in standards across cities, the existing models based on static image training are difficult to capture the dynamics of subjective perception. In this paper, we design an LLM-Based Intelligent Agent for Subjective Perception of Urban Streets (SPUS-Agent), which includes two operational stages: Multidimensional subjective perception benchmarking and subjective perception prediction. Among them, in the subjective perception benchmarking stage, the SPUS-Agent automatically collects street view training images on an electronic map and employs a comparative learning approach using a large model to obtain the benchmark. In the subjective perception prediction stage, the SPUS-Agent can move to any location in the city and obtain the subjective perceptual score of the location by collecting the current street view images and comparing them with the perception benchmark. Experimental validation demonstrates the SPUS-Agent’s significant predictive capability (accuracy: 92.4
The Budgeted Influence Maximization (BIM) problem aims to maximize social influence under cost constraints, yet existing methods struggle with computational complexity and adaptability to diverse cost models. We propose ComIVY, a novel approach integrating community structure detection with the IVY algorithm which is a new bio-inspired modeling variant drawing inspiration from the growth patterns of Ivy plants. ComIVY dynamically allocates budgets across communities, selects seed nodes via IVY optimization, and reduces influence overlap through a two-hop neighbor filtering mechanism. We evaluate ComIVY on six real-world networks, demonstrating up to 25
Temporal knowledge graphs (TKGs) with evolving event representations serve as critical infrastructure for AI applications like information retrieval and rumor detection. Current TKG completion methods face three key limitations: (1) Conversion of quadruple structures to triplets constrains temporal information representation, (2) Cross-space interactions among entities, relations, and timestamps are challenging, and (3) Neighborhood contexts are underutilized and disambiguation mechanisms for polysemic elements are absent. To address these challenges, we propose the Quadruple-based Neural Discourse-Aware (Q-NDA) network. Our framework employs space-specific encoders for entity, relation, and timestamp representations, coupled with an Information Discourse Module enabling bidirectional cross-space communication. The architecture incorporates contextual disambiguation and neighborhood attention mechanisms to resolve semantic ambiguity while aggregating structural patterns from adjacent quadruples. This constitutes the first systematic approach addressing completion in TKGs through integrated neighborhood exploitation. Evaluations on benchmark datasets demonstrate Q-NDA’s superior performance, particularly achieving state-of-the-art results on the geometrically complex GDELT dataset - yielding 19.3
Representation learning is key in trajectory analysis, affecting downstream task performance. To reduce computational overhead from numerous trajectory points, spatial partitioning of the road network is often used. Existing spatial partitioning methods primarily include grid-based partitioning and road network partitioning, which aim to achieve a more uniform distribution of trajectory points. However, after partitioning, these methods typically employ one-hot encoding to represent subregions, failing to capture the spatial proximity relationships between adjacent subregions. Moreover, due to variations in road network structures, existing trajectory representation learning methods often require complete retraining when adapting to new maps, resulting in high computational costs. This paper proposes a Partition Transition Probability and Pretraining Graph Attention Network (PTP-PGAT). The model integrates a road network partitioning strategy and incorporates a partition transition probability mechanism to capture spatial proximity between sub-regions effectively. Additionally, leveraging the hierarchical partition tree structure of road network partitioning, PTP-PGAT supports a pretraining-finetuning transfer learning paradigm. After pretraining on one dataset, the model can efficiently adapt to new datasets with minimal fine-tuning. Experiments show PTP-PGAT outperforms baselines, improving accuracy by up to 15.5
Entity alignment, which identifies equivalent entities across source and target knowledge graphs, is essential for constructing comprehensive knowledge bases. Although TransE-based methods show effectiveness, they struggle to capture complex relations due to limited embedding expressiveness. The advanced TransD model uses dynamic mapping matrices but remains constrained by static relation modeling, semantic ambiguity, and insufficient generalization for low-frequency relations. To address these challenges, this paper proposes a Multi-Granularity Dynamic Projection Entity Alignment (MG-DPEA) model by integrating fine-grained and coarse-grained relation projections. Firstly, the fine-grained layer retains TransD’s entity relation collaborative projection to capture specific relational features. Secondly, the coarse-grained layer uses hierarchical semantic clustering based on WordNet and the Stick-Breaking algorithm, grouping semantically similar relations that share projection matrices. Finally, an adaptive fusion unit dynamically integrates these multi granularity knowledge graphs. Extensive experiments demonstrate that MG-DPEA significantly improves alignment performance, especially for low-frequency relations, providing an interpretative and efficient approach for knowledge graph fusion under resource-constrained scenarios.
