| 摘要: | 環保犯罪偵測對環境保護來說是一個重要議題，涉及任意棄置和掩埋廢 棄物所引發的環境污染問題，以及利用不實申報廢棄物賺取非法利益等。過去常使用基本的質量平衡、統計資訊分析或機器學習來辨別可疑資料，以判定是否存在不實申報。然而，若採用質量平衡和數值報表等數學統計方法進 行資料分析與檢驗，需要針對目標廠商進行數據檢查，可能會遺漏潛在的可疑事業群。此外，一旦廠商被判刑或處以罰鍰，他們未來很可能會造假資料以躲避質量平衡的稽查。而機器學習模型訓練則常忽略時間訊息，並可能缺乏上下游關係的考量。
為了解決上述問題，我們在本篇論文中提出一種使用圖形卷積網路(GCNs)進行環境異常檢測的新方法。在實驗中，我們提出兩種圖形卷積網路模型的架構，並進一步比較兩者的效能，分析影響預測結果的可能要素。本篇論文首次導入深度學習來構建環保資料的上下游關係，並保留時間資訊，以進行環保犯罪偵測。我們的主要貢獻如下:首先，我們是首個利用GCNs檢測環境資料異常的研究;其次，我們使用無向鄰居傳播演算法將複雜的資料建構成圖;第三，我們進行了包含和不包含時間因素的 GCNs的對比分析。;Detecting environmental crimes is a crucial issue for environmental protection, involving problems related to illegal disposal and burial of waste, resulting in environmental pollution, as well as earning illegal profits through false reporting of waste. Traditionally, basic mass balance, statistical information analysis, or machine learning have been used to identify suspicious data to determine whether false reporting exists. However, using mathematical and statistical methods like mass balance and numerical reports for data analysis and inspection requires checking data for target enterprises, which may miss potentially suspicious enterprises. Furthermore, once an enterprise is convicted or fined, they are likely to falsify data in the future to evade mass balance inspections. Machine learning model training often ignores temporal information and may lack consideration of upstream and downstream relationships.
To address these issues, in this paper, we presented a novel approach for environmental anomaly detection using Graph Convolutional Networks (GCNs). In our experiments, we propose the architectures of two graph convolutional network models, compare their performance, and analyze the possible factors influencing prediction results. This paper is the first to introduce deep learning to construct upstream and downstream relationships in environmental data, retaining temporal information for environmental crime detection. Our main contributions are the following: firstly, we are the pioneers in utilizing GCNs for detecting anomalies in environmental data; secondly, we constructed complex data into a graph using the Undirected Neighbor Propagation algorithm; and thirdly, we conducted a comparative analysis between GCNs with and without temporal considerations. |
