基於時空圖神經網絡之海上犯罪預測

Abstract

犯罪預測在近幾年逐漸成為一個重要的議題，其有助於協助政府維護社會安全。而隨著深度學習的成熟和圖神經網絡的興起，越來越多學者投入這個領域的研究，透過各種方法來增加預測的準確率。然而，目前的研究都僅限於陸地上的犯罪預測，忽略的預防海上犯罪的重要性。海上犯罪影響的層面涉及社會治安及環境保護，且擁有的人力資源更稀少、範圍更廣，因此海上犯罪預測任務成為一大挑戰。

在本研究中，我們提出了一個名為ST-AAGAT的新時空圖網絡模型架構。此模型致力於透過圖神經網絡找到空間的相關性，再藉由Autoformer架構尋找犯罪發生的週期性，並結合捕捉到的時間和空間特徵進行時間序列的預測。我們將此模型在台灣海巡署的真實犯罪資料集上進行多個不同面向的實驗，結果顯示 ST-AAGAT相較於其他模型有更好的表現。透過這項研究，我們希望能夠協助海巡署更有效地進行人力派遣，同時實現海上犯罪的預防和提高環境保護的意識。

Crime prediction has become an important issue in recent years as it helps governments maintain social security. With the maturity of deep learning and the rise of graph neural networks, more researchers have been studying this field, employing various methods to enhance prediction accuracy. However, current research is limited to crime prediction on land, neglecting the significance of preventing maritime crimes. Due to the limited human resources and a broader scope, maritime crime prediction has become challenging.

In this study, we propose a novel spatial-temporal graph neural network model called ST-AAGAT. This model aims to identify spatial correlations through graph neural networks and leverage the Autoformer architecture to find the periodicity of crime. After that, we combine temporal and spatial patterns to predict the crime occurrence probability from time series data. We conducted experiments on maritime crime datasets with various aspects, and the results demonstrated that ST-AAGAT outperformed other models. Through this research, we hope to assist the coast guard administration in deploying resources more efficiently, preventing maritime crimes, and raising awareness for environmental protection.