資料探勘演算法於軍人貪污量刑之預測及比較

Sentence Prediction and Comparison of Data Exploration Algorithms for Corruption Conviction in the Military

本研究為預測軍人貪污犯罪之量刑，藉以驗證資料探勘演算法可否運用法律刑度判決的可行性，提供未來軍人涉犯貪污犯罪之智慧量刑預測系統開發之基礎。因此，本研究透過所蒐集到之425件軍人貪污判決書，萃取、歸納出27項因素後，透過關聯規則進行因素縮減，再將樣本區分為訓練組與測試組，並針對訓練組以增加少數抽樣法解決不均衡資料問題後，以類神經網路與決策樹等資料探勘演算法進行量刑預測，並依據預測結果進行正確率比較，經研究得到以下結果：一、類神經網路模型測試結果具79.69％之正確率，而決策樹模型則高達90.62％之正確率。由此結果得知在針對軍人貪污量刑之預測上，決策樹模型具有較佳的預測結果。二、資料探勘分類預測之投入項目愈少，其運算效能與速度愈快。本研究最後各自以類神經網路與決策樹所得之十項重要因素作為投入因素再次進行分析，驗證模型準確度仍然能維持預測正確率。本研究首先嘗試以軍人貪污量刑為標的抛磚引玉，透過軍人貪污犯罪判決書構成之三大類之屬性進行投入因素之萃取，再以類神經網路與決策樹演算法進行量刑模型建構與預測模型之驗證，發現決策樹演算法表現較佳。最後再次進行投入因素之縮減，發現各演算法僅需十項投入因素即可獲得相同之預測正確率。未來，後續研究者可再嘗試其餘演算法，或可將此系統實作軍人涉犯貪污犯罪之智慧量刑預測系統之開發。
This study aims to predict the sentencing of corruption conviction committed by military personnel. In order to see the feasibility of predicting sentencing, the study adopts the data exploration algorithm method and lay the foundation for future development of smart sentencing prediction system that is targeted at the military personnel committing corruption. Hence, 27 input factors concluded through 425 corruption judgement cases are generated. To reduce the number of input factors by association rules, sample factors are divided into training and test group. To prevent it from compromising with insufficient information, the study increases the small fraction of sampling in the training group, the test of results using neural network and decision trees model is concluded as follows, 1. The overall prediction rate of the neural network model is 79.69%, and the decision trees model is as high as 90.62%. As the results presented, decision trees model is better to predict the sentencing of corruption offense committed by military personnel. 2. Fewer input factors in the Data Exploration Classification reflect a better algorithm and faster speed. This study concludes 10 vital factors through neural network and decision tress models. By analyzing these 10 factors, the results are no difference. This study uses corruption conviction cases committed by military personnel as a pioneer approach. Through the judgement in which is comprised of 3 parts, a number of input factors is extracted. A test run by implementing those factors in neural network and decision tress models indicates decision trees model produces the best results. Furthermore, by reduction of number of input factors, the study shows with only 10 input factors the results remain the same. In the future, follow-up researchers might adopt other algorithms implementing for development of smart sentencing prediction system that is targeted at the military personnel committing corruption.

This study aims to predict the sentencing of corruption conviction committed by military personnel. In order to see the feasibility of predicting sentencing, the study adopts the data exploration algorithm method and lay the foundation for future development of smart sentencing prediction system that is targeted at the military personnel committing corruption. Hence, 27 input factors concluded through 425 corruption judgement cases are generated. To reduce the number of input factors by association rules, sample factors are divided into training and test group. To prevent it from compromising with insufficient information, the study increases the small fraction of sampling in the training group, the test of results using neural network and decision trees model is concluded as follows, 1. The overall prediction rate of the neural network model is 79.69%, and the decision trees model is as high as 90.62%. As the results presented, decision trees model is better to predict the sentencing of corruption offense committed by military personnel. 2. Fewer input factors in the Data Exploration Classification reflect a better algorithm and faster speed. This study concludes 10 vital factors through neural network and decision tress models. By analyzing these 10 factors, the results are no difference. This study uses corruption conviction cases committed by military personnel as a pioneer approach. Through the judgement in which is comprised of 3 parts, a number of input factors is extracted. A test run by implementing those factors in neural network and decision tress models indicates decision trees model produces the best results. Furthermore, by reduction of number of input factors, the study shows with only 10 input factors the results remain the same. In the future, follow-up researchers might adopt other algorithms implementing for development of smart sentencing prediction system that is targeted at the military personnel committing corruption.