本研究旨在建構一套融合結構屬性、空間資訊與心理參考價格之深度學習房價預測模型，並以新北市板橋區為實證場域，探討其於價格合理性判定上的應用潛力。針對實價登錄資料中可能存在的價格異常問題，本研究提出以多層感知神經網路（Multi-Layer Perceptron, MLP）為基礎架構，並結合殘差學習模組（Residual Block）與多尺度空間異常值剔除機制（MSSOD），以提升模型的訓練穩定性與非線性特徵擬合能力。
研究資料涵蓋 108 至 113 年間板橋區逾三萬五千筆住宅交易紀錄，經資料清理、異常值檢測、地理變數建構與特徵工程後，建立可供模型訓練之輸入特徵。預測目標設定為實際價格與鄰近加權均價之對數差值（log_price_diff），並以 ±10% 誤差區間為合理價格判定標準，設計 ACC@10 作為核心評估指標。
實證結果顯示，本研究模型於測試資料集中達成 ACC@10 為 70.17%、MAPE 為 11.49%、RMSE 約為 22,914 元，顯示具備良好的準確性與實務可用性。進一步以 30 次獨立訓練檢驗穩定性，平均 ACC@10 為 68.95%，標準差僅 0.56%，驗證模型具備高度可重現性。價格區段分析結果亦顯示，模型於中價樣本表現最佳，而高低價樣本之預測誤差相對較大。組件消融實驗則證實，殘差結構與異常值剔除機制皆為提升效能之關鍵要素。
整體而言，本研究模型在理論設計與實務應用層面均具貢獻，能作為市場價格合理性判斷、異常交易預警及政策監理輔助工具之基礎架構。

This study proposes a deep learning model for housing price prediction that integrates structural attributes, spatial information, and psychological reference prices, using Banqiao District, New Taipei City as a case study. The model adopts a Multi-Layer Perceptron (MLP) enhanced with residual blocks and a multi-scale spatial outlier detection (MSSOD) mechanism to improve training stability and nonlinear fitting.
Using over 35,000 residential transactions from 2019 to 2024, the prediction target is defined as the log difference between actual prices and nearby weighted averages. A ±10% deviation is set as the threshold for rational price evaluation, with ACC@10 as the main metric.
Results show the model achieves an ACC@10 of 70.17%, a MAPE of 11.49%, and an RMSE of about 22,914 NT dollars on the test set. Thirty independent runs confirm robustness, with an average ACC@10 of 68.95% and low variance. Ablation studies further demonstrate that both residual learning and spatial filtering are critical to performance.
Overall, the proposed framework provides practical value for rational price assessment, anomaly detection, and housing market supervision.

1.Anselin, L. (1995). Local indicators of spatial association—LISA. Geographical Analysis, 27(2), 93–115.
2.Chen, J., Gao, J., Wu, X., & Xie, L. (2020). House price valuation model based on geographically neural network weighted regression: The case study of Shenzhen, China. Sustainability, 12(16), 6674.
3.Cheng, T., & Li, Z. (2004). A multiscale approach to detect spatial-temporal outliers. Proceedings of the 2004 ACM Symposium on Applied Computing, 576–583.
4.Fotheringham, A. S., Brunsdon, C., & Charlton, M. (2002). Geographically weighted regression: The analysis of spatially varying relationships. Chichester: Wiley.
5.Getis, A., & Ord, J. K. (1995). Local spatial autocorrelation statistics: Distributional issues and an application. Geographical Analysis, 27(4), 286–306.
6.He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770–778.
7.Huang, B., Wu, B., & Barry, M. (2010). Geographically and temporally weighted regression for modeling spatio-temporal variation in house prices. International Journal of Geographical Information Science, 24(3), 383–401.
8.Karkar, A., Shekhar, S., Fei, X., & Adams, B. (2020). A multiscale approach for spatio-temporal outlier detection. GeoInformatica, 24(1), 25–63.
9.Kuchler, T., Piazzesi, M., & Stroebel, J. (2023). Housing market expectations. The Quarterly Journal of Economics, 138(1), 1–49.
10.Lu, W., & Chen, Y. (2007). A multi-scale approach to detect spatial outliers. Computers, Environment and Urban Systems, 31(6), 567–587.
11.Selim, H. (2009). Determinants of house prices in Turkey: Hedonic regression versus quantile regression. Applied Economics Letters, 16(17), 1705–1710.