邊境人工智慧預測系統──以羅吉斯迴歸評估類別型因子不同編碼方法之比較   全文下載

Border Prediction Intelligent System-Evaluating Different Encoding Methods for Categorical Factors by Logistic Regression

為提升邊境輸入食品安全，透過衛生福利部食品藥物管理署食品雲及國內外開放資料，結合數據科學及機器學習演算法，導入邊境人工智慧預測系統，運用人工智慧模型進行決策輔助。本研究檢討現行系統對於類別型因子之編碼方式，並探討其他常用之編碼方法，以評估不同編碼方法對模型之影響。根據分析結果，現行使用類別型因子之編碼方式，呈現出較高敏感度及較低精確度之特性，而本研究使用目標編碼方式建構之模型，則呈現出較低敏感度及較高精確度之特性，惟在使用測試集資料進行驗證時，高敏感度特性的效果不明顯，使得目標編碼方式建構之模型表現較現行模型更佳，建議後續可評估於邊境人工智慧預測系統中之類別型因子使用目標編碼方式進行前處理，以期優化邊境人工智慧預測系統之預測能力，提升邊境輸入食品之安全。

To promote the safety of imported food at the border, a border prediction intelligent system has been implemented by leveraging the Food Cloud of TFDA and open data from domestic and international sources, combining data science and machine learning, and using artificial intelligence models for decision support. This study reviewed the encoding methods for categorical factors in the current system and explored other commonly used encoding methods to evaluate their impact on the modeling. According to the analysis results, the current encoding method for categorical factors exhibited higher sensitivity and lower precision. In contrast, the model constructed using target encoding method in this study demonstrated lower sensitivity and higher precision. However, while validating with test data set, the effectiveness of high sensitivity characteristics was not significant, so that the model constructed using target encoding method outperformed the current one. It is recommended to evaluate the models using target encoding method for categorical factors in the border prediction intelligent system, in order to optimize its predictive capabilities and ensure the safety of imported food at the border.