以交通特徵預測台北市房地產地價之研究-應用類神經網路之研究途徑

Abstract

台北市的房價居高不下，時至今日，要在台北市買一棟房子對於年輕人而言已非易事。在房地產市場中有許多因素都會影響房地產物件價值，不僅是房地產市場景氣，鄰里環境、生活機能、交通便利等等因素更是有顯著影響力的變因，而在討論台北市房價時，房地產具有多少價值才是合理價格是許多人心中的疑問，也是本研究主要探討之議題。\r\n\r\n在眾多特徵值中，過去的研究中證實土地及房屋的可及性與其至市中心的距離大多主要是影響房地價的條件，Agostini and Palmucci (2008)亦提到大眾交通運輸工具帶來的交通易及性不僅使得生活便利，更帶來房地產價值的提升。基於過去基礎，本研究以路網完整性的結構作為影響變數之選擇，包含城市間移動的捷運系統，範圍較小但較密集的公車網路，以及在人們日常生活圈移動中做為代步工具的公共自行車(YouBike)。\r\n\r\n本研究將就路網系統的面向，採用民國102-105年台北市實價登錄資料為樣本，透過將台北市網格化並依照每年度做時間調整，將台北市分為多個邊長100公尺的方形網格，經文獻探討後，採用交通特徵捷運、公車及公共自行車共3個變數作為模型輸入，並利用倒傳遞類神經網路方法建立地價預測模型，以單位網格作為輸入樣本訓練預測模型，探討交通特徵與地價的關聯性。\r\n\r\n本文主要分為兩個研究階段，第一階段主要分為兩組實驗，預測各種交通特徵組合所代表的平均地價等級研究。第一組是將民國102年至民國105年總計共37840個單位網格樣本，隨機選取70%作為訓練集、30%作為測試集，分別做將地價分為四、五、六個等級的實驗，透過預測地價等級模型的預測結果分別為80.21%、79.81%、79.65%；而第二組是進階的將台北市民國102年至民國105年的單位網格合併成共2365個邊長400公尺大型網格，在這些大型網格內，依序從中機選出70%的單位網格作為訓練集、30%的單位網格作為測試集，同樣分別做將地價分為四、五、六個等級的實驗，透過預測地價等級模型的預測結果皆為79.26%。\r\n\r\n第二階段研究主要是利用民國102年至民國105年資料訓練的地價預測模型，用來預測民國106年的平均地價並驗證之，此階段研究將平均地價區分為高價地段及低價地段區域，預測各網格未來之地價等級，所得平均準確度為65.71%。\r\n\r\n透過此模型不僅可得各網格內所具有的特徵條件，亦能透過交通特徵組合預測各網格應具有之合理地價等級，並透過現有資料預測未來各網格分屬高價區域或低價區域。未來能增加特徵值或改用深度學習模型來進行試驗，加強地價預測模型準確度，亦能朝預測網格漲跌趨勢研究以增進此議題的研究廣度及深度。

Because of the unusual high price of real estate in Taipei City, young mans have difficulties to buy a house for themselves. The problems deteriorate these days, so the research propose a model to predict the price of real estate in Taipei City.\r\n\r\nAmong several features that can affect the price of real estate, the transportation features play the main role. In the past, researches also proved that the transportation accessibility has a great impact on real estate values. Based on these research, we choose three factors in our model including Taipei Metro that can extend the accessibility between cities, the bus lines that serves the entire city and You-Bike that covers smaller areas without bus services.\r\n\r\nAccording to those issues mentioned above, the purpose of this research is to estimate the landing price in Taipei City by using public transportation factors. Applying a data set with 86,716 real transaction records of real estate in Taipei in 2013-2016, we cut a Taipei City map into multiple grids, then we map the transactions to these grids. Based on these grids of side 100 meters, using a Back-Propagation Neural Network (BPN) to estimate the landing price in Taipei City. This research would like to know the relationship between traffic features and landing prices.\r\n\r\nThe research will conduct two experiments to estimate the landing price in Taipei City by the traffic feature combinations, 37840 grid data will be used. We select 3 variables which may impact the value of real estate as input. Empirical study shows that: Firstly, we will pick 70% of the grid data in a random way as training data and the others as testing data. Dividing the landing price to the fourth level, fifth level, and sixth level, and observing the estimation performance. The result reveals that: 1) In the fourth level, the estimation accuracy is 80.21%. 2) In the fifth level, the estimation accuracy is 79.81%. 3) In the sixth level, the estimation accuracy is 79.65%. Secondly, we will merge these grids of side 100 meters into grids of side 400 meters. Then we pick 70% of the grid data in a random way from each grid of side 400 meters as training data, the others as testing data. Similarly, we grade the landing price to the fourth level, fifth level, and sixth level. The result reveals that no matter what level it is, the result is 79.26%. In the future, adding more features or using deep learning model may be a good way to improve the model performance. This estimation model can also be a reference of the support system of landing price valuation in the future.