運用羅吉斯迴歸與倒傳遞網路進行二元類別選擇之比較——對住宅成交機率之測試實驗

The Comparison between Logistic Regression and Back-propagation Network on Binary Choice-a Test of Housing Market Transaction Data

探討住宅成交與否二元類別選擇之統計方法，一般包括羅吉斯迴歸(1ogistic regression)、普羅比迴歸(probit regression)與區別分析(discriminate analysis)，其中以羅吉斯迥歸最常被使用。此外，類神經網路之倒傳遞網路(back-propagation network)、機率神經網路(probabilistic network)以及輻射式函數網路(radial basis function network)亦可應用於二元類別選擇分析，其中以倒傳遞網路最常被使用。相關研究多以正確預測率，作為羅吉斯迥歸與倒傳遞網路模式優劣之判斷準則。然而，本研究認為羅吉斯迴歸之最低正確預測率，受二元類別資料之原始樣本比例所影響。而當樣本比例懸殊時，倒傳遞網路亦可能發生過度學習某類型樣本，而無法辨別另一類型樣本的問題。因此正確預測率的高低，可能無法作為模式優劣之判斷指標。本研究以台南市東區之住宅交易資料進行實證分析，以等比例以及原始樣本比例分別建構模式。藉以瞭解不同樣本比例，對羅吉斯迴歸與倒傳遞網路正確預測率之影響。

When we analyze the behavior of binary choice, the appropriated econometric methodologies include logistic regression, probit regression and discriminate analysis, and logistic regression is most frequently used. Artificial neural network can also be applied to the binary choice, and back-propagation network (BPN) is most frequently used. We employ logistic regression and back-propagation network to analyze this question, and focus on the predictive performance of the two models. Previous studies used the accuracy rate of prediction as a comparison criterion between logistic regression and back-propagation network. However, the lowest prediction accuracy rate of logistic regression is affected by the sample ratio, we cannot use the accuracy rate to compare the performance of the two different models. In this paper, we will discuss the advantages and disadvantages of logistic regression and BPN, and compare the predictive performance on the same data set. Furthermore, we eliminate some data to get an equal sample ratio data set, and analyze the effect of sample ratio on predictive performance by actual sample ratio and equal sample ratio data set.