以機器學習方法研究臺灣各大縣市影響房地產之特徵

Abstract

在台灣地區的學術界鮮少使用較新的梯度提升決策樹（GBDT）針對房地產鑑價的應用進行研究，此外，針對不同的變數類別在各大縣市的貢獻度比較則更少。本篇論文採納國內外針對房地產特徵變數的建議，加入了地理位置、時間刻度、總體經濟指標、生活圈、空氣污染與線上行為的變數類別於模型之中，並使用輕量級梯度提升機器（lightGBM）和SHAP的分析套件來展開研究方法。本論文的目的是逐一的評估各大變數類別在全台灣七大主要城市的貢獻度表現，並同時檢測資料集的“時效性”存在與否與對預測結果產生的影響。研究結果發現，地理區位和生活圈的變數可獨立存在為模型主要參考依據，而資料的時效性充斥著整個資料集，建議未來房地產的相關應用應盡可能的更新訓練資料集，或在資產重新估價的專案上，挑選訓練資料時應盡可能包含欲估價的時間範疇。

There are few studies in Taiwan using latest GBDT models on real estate market and are even less of them making feature comparison across individual cities. Draw from the studies of previous scholars in Taiwan and oversea, we included feature categories from location, time, economy, living area, pollution and online behavior and utilized the related analysis toolkit of lightGBM and SHAP value. The purpose of this study is to evaluate the contribution of different categories on major cities in Taiwan and also to examine the timeliness of dataset when training a model. The result turned out to be location and living area standalone are most contributing categories. And, data timeliness issue is all over the dataset, it is suggested to train a most up-to-date possible model for predicting the future or covering the re-appraisal period when choosing the training set.