A graph neural network-enhanced knowledge graph framework for intelligent analysis of policing cases.

Area of Science:

• Artificial Intelligence
• Computer Science
• Data Science

Background:

• Traditional convolutional neural networks (CNNs) have limitations in processing complex, multi-feature data.
• Existing knowledge graph embeddings may not fully capture the nuances of policing case data.

Purpose of the Study:

• To develop a novel graph neural network (GNN)-enhanced knowledge graph framework for policing case prediction.
• To improve prediction accuracy and model efficiency by integrating graph structures with deep learning.

Main Methods:

• Constructed a knowledge graph for policing cases using graph neural networks.
• Employed Label Propagation Algorithm (LPA) with Convolutional Graph Networks (GCN) for edge weight training.
• Enhanced traditional CNNs into a multichannel network to process multiple policing case features.

Main Results:

• Achieved a prediction accuracy of 87.7% using the multichannel CNN approach.
• Integrated an efficient pairwise feature extraction module to enhance network backbone.
• Demonstrated significant reductions in computational complexity (53.5% fewer FLOPs) and parameters (70.2% fewer).

Conclusions:

• The proposed GNN-enhanced knowledge graph framework effectively improves policing case prediction accuracy.
• The multichannel CNN architecture accommodates diverse feature factors, expanding perceptual fields for better predictions.
• The method offers a superior balance between predictive performance and computational efficiency compared to existing work.