運用貝氏分類與類神經網路於金融業授信不正常繳款之研究

Application of Na?ve Bayes Classifier and Artificial Neutral Networks to Abnormal Credit Payment in Finance Industry

趙伯鈞(Po-Chun Chao)

指導教授 ： 盧以詮

元智大學/資訊學院/資訊管理學系/碩士(2011年)

摘要

在資料探勘的模型中，類神經網路為人所詬病的原因在於模型的產生為黑盒子，無法由統計數學算式取得明確的結果，所以本研究利用貝氏分類容易計算取得模型結果，若屬性間為近似條件獨立下，可以得到良好的分類效果的原因下與類神經網路比較。因研究資料的欄位屬性相依性很高，所以本研究希望在設定條件相等下，確認類神經網路的訓練模型結果優於貝氏分類，再以不同整類神經網路方法產生不同的訓練模型結果，取其較佳的類神經網路方法後，加以欄位屬性的離散化處理，使最後訓練的類神經網路模型為最佳的模型，並依此訓練建立的模型可做為金融業授信放款時?考之依據，進而給予適當的授信放款額度，進而拒絕授信放款，使風險發生時降低銀行的損失。

關鍵字

資料探勘 ； 貝氏分類 ； 類神經網路

並列摘要

Among the models of Data Mining, one of the big problems that artificial neural networks has is that the production of the models is like a black box, and that it is difficult to get definite results through statistically mathematic equation. To solve this problem, the study in this thesis employs Na?ve Bayes Classifier, which is easy to calculate, to get the model results and to get better classification effects in comparison between Na?ve Bayes Classifier and Artificial Neutral Networks if the attributes are similar and independent. Because the compatibility of attributes of the research data is very high, the study hopes to make sure the training model results of Artificial Neural Networks are better than those of Na?ve Bayes Classifier and to use different Artificial Neural Networks methods to produce different training model results with the equally set conditions. Besides, the study uses Discretization of attributes to make the last trained Artificial Neural Networks the best model after getting the better Artificial Neural Networks methods. Finance industry can use the trained models resulted from the study as a reference when giving credits and making loans and thereby can give adequate credit and loan lines or can refuse to give credits or make loans to reduce their loss when risks arise.

並列關鍵字

Data Mining ； Na?ve Bayes Classifier ； Artificial Neural Networks

參考文獻

楊金聲，「利用類神經網路與線性迴歸進行成本預測之研究—以印刷電路板產業為例」，中原大學資訊管理學系，碩士論文，2006

Bacardit, J. and Garrell, J., “ Evolving multiple discretizations with adaptive intervals for a Pittsburgh Rule-Based Learning Classifier System,” Genetic and Evolutionary Computation Conference (GECCO03) , pp.1818-1831, 2003.

Boulle M., “Khiops: A Statistical Discretization Method of Continuous Attributes,” Machine Learning, vol.55 , pp.53-69, 2004.

Chan, C., Batur, C., and Srinivasan, A., “Determination of quantization intervals in rule based model for dynamic,” In Proceedings of the IEEE Conference on Systems, Man, and Cybernetics, pp.1719–1723, 1991.

Elomaa, T. and Rousu, J., “Efficient multisplitting revisited: Optima-preserving elimination of partition candidates,” Data Mining and Knowledge Discovery, Vol.8, pp.97-126, 2004.