透過您的圖書館登入
IP:34.230.84.106
  • 學位論文

實價登錄之類神經網路估價模型-以高雄市農16及美術館區大樓為例

Appraisal Model of Artificial Neural Network based on Real price System in Lon 16 and Museum in Kaohsiung

指導教授 : 紀雲曜

摘要


國外自1960年代開始應用電腦輔助大量估價(CAMA),做為稅務上之輔助工具。電腦輔助大量估價是利用電腦輔助提供標準化程序,來進行大量不動產估價。一般應用於大量估價模型工具包含複迴歸、類神經網路、時間序列分析及稅賦評價模型等;其中,類神經網路為一條新穎且具前瞻性方法。 2012年8月國內正式實施實價登錄,公告不動產交易資訊。為進一步瞭解類神經網路對應實價登錄資訊時,是否能具高效益之預測能力,本研究嘗試以監督式倒傳遞類神經網路(以下簡稱倒傳遞網路)來預測實價登錄下的不動產交易總價。 在應用上,倒傳遞網路目前遇到的最大困擾在於網路參數的設定,不同的輸出入變數結構及資料分布差異將影響網路學習效果,須經由不斷地試誤以尋求最適模式。由研究成果顯示影響倒傳遞網路預測能力之參數,包含輸入層變數數目、第1層隱藏層處理單元數目、第2層隱藏層處理單元數目、訓練及測試範例比例及網路學習循環數目等參數。 透過類神經網路分析結果之敏感度分析得到影響網路預測結果之主要輸入變數,分別為建物面積、屋齡及總樓層等三個變數。若加入其他變數試圖改善解答品質,其改善幅度並不大。經實證成果顯示,倒傳遞網路於不動產實價登錄交易總價之預測能力符合預期研究設定。平均絕對誤差百分比介於9.48~13.92%之間,低於設定值20%之要求;絕對誤差百分比小於10%命中率介於45.6%~56.7%之間,高於設定值30%之要求;平均絕對誤差百分比小於20%命中率介於76.5%~83.3%之間,高於設定值70%之要求。 另外,農16和美術館合併後的樣本,預測效益高於各別的單獨區域,顯示在類神經網路模型下,大樣本的預測能力較佳。而絕對誤差大於30%有52.9%來自高總價樣本,這也顯示網路對於高交易總價的預測能力較低。從研究結果比較得知,類神經網路模型下的MAPE值為9.48%優於特徵價格法的17.1%。

並列摘要


The computer assistance mass assessment (CAMA) was used overseas to estimate the tax affairs from 1960. The computer assistance mass assessment provides the standardized procedure by using the computer assistance, carries on the massive real estate estimate. Applies generally in the massive estimate model tool contains the duplicate regression, neural network, the time series analysis and the taxes appraisal model and so on; Among them, the neural network is a forward-looking method. In August, 2012 the real price registration system officially started, announcement real estate transaction information. For finding out if the neural network has high forecast ability correspondence actual price registers, we try to use back-propagation neural network (BPN) to estimate the dealing price from actual price registration system. In application, the biggest problems of the back-propagation network is the set of network parameters, a different structure and data input and output variables will affect the distribution of differences in network learning effect, the optimum mode needs to be found by the try-and-error. By the research results , the parameters affectd the BPN are input variable number and the first hide level processing unit number and the second hide level processing unit number, the ratio of the training and the testing samples and the circulation number. Using the sensitivity analysis getting from the results of the neural network, we foung the major input variables affect network prediction are building area, age and total floors. Adding other variables try to improve the quality of answers, the magnitude of the improvement is not significant. From the result we found that the forecast ability of BNP using in the actual price registration system to conform to the anticipated research hypothesis. The mean absolute percentage error (MAPE) between 9.48%~13.92% is lower than the setting value 20%; The hit-rate of absolute error below 10% is between 45.6%~56.7%, which is higher than the setting value 30%; The hit-rate of absolute error below 20% is between 76.5%~83.3%, which is higher than the setting value 70%. In addition, the predict effective in combined samples of Lon 16 and museum is better than the separated area, it means the predict ability of a large sample is better in artificial neural network model. There are 52.9% which absolute error higher than 30% from the high price samples, it shows the predictive ability of high total price is lower in neural network model. From comparative studies, MAPE value of the neural network model is 9.48% better than 17.1% of the hedonic price method.

參考文獻


24. 光特資訊科技股份有限公司(2013)。內政部不動產交易實價查詢服務網系統PC(網頁)版系統操作手冊。
13. 林英彥(2004)。不動產估價。台北市:文笙。
17. 陳奉瑤、楊依臻(2007)。個別估價與大量估價之準確性分析。住宅學報,16(2),67-85。
16. 賴碧瑩(2007)。應用類神經網路於電腦輔助大量估價之研究。住宅學報,16(2),43-65。
18. 龔永香、江穎慧、張金鶚(2007)。客觀標準化不動產估價之可行性分析-市場比較法應用於大量估價。住宅學報,16(2),23-42。

被引用紀錄


黃威樺(2015)。高中生運動社團參與動機、升學壓力與課業延宕關係之研究〔碩士論文,淡江大學〕。華藝線上圖書館。https://doi.org/10.6846/TKU.2015.00834
陳坤厚(2012)。國中生運動性社團參與動機、人際關係與學習成效之研究〔碩士論文,朝陽科技大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0078-1511201214172662

延伸閱讀