應用文字探勘技術於住宅房價模型──以新北市汐止區為例

購屋是人生中重大的決策。然而除了房屋的基本資料如：鄰近物件的實價登錄資料及房屋特徵資訊外，尋求其他資訊來源或是區域購屋經驗分享，都是輔助購屋人做出決策的重要步驟。本研究以新北市汐止區為例，以不動產實價交易資料以及批踢踢實業坊的房價討論版作為資料來源。研究首先建立含文字探勘結果的特徵價格模型，藉此分析影響房價的變數。接著，將討論版資料、每坪單價、成交筆數視為時序變數，建構房價預測模型。結果顯示，基礎特徵價格模型在加入文字探勘資料後，配適獲得改善。在房價預測模式中，以預售屋模型表現最佳，考慮文字探勘資料時可提升3.5%的模型解釋能力，該模型的平均絕對百分比誤差僅有18.12%。研究同時發現，討論版文章數對住宅價格有正向影響，尤其是在預售屋模型中其影響更為明顯。本研究顯示傳統房價預測模式與文字探勘結果結合的可能性，此模式可以推廣至不同區域，藉此提供在購屋過程中相對弱勢的買方更正確的參考資料。

Buying a house is an important decision yet further information besides pricing and house information is needed. The study takes Xizhi Distict as an example. We collect the real estate actual selling price registration data from Department of Land Administration and articles from the most famous bullinton board system (BBS) in Taiwan. We first build augmented hedonic price models with collected data to analyze the factors that affect housing price. After that, we treat articles and comments from BBS, price per unit, and the mumber of transaction as time series parameters to build prediction models. The results show that the augmented models have better fit than the hedonic price models. The prediction model for pre-sale house shows a 3.5% improvement by considering time series parameters, and the mean absolute percent error reaches 18.12%. The study also finds the number of articles on BBS will influence the housing price positively especially in the pre-sale house model. This research demonstrates hedonic price models with textmining results provides more accurate predictions that helps individual sellers and buyers suffer from information insufficiency.