This paper presents REALM, an open-source tool for utilizing machine learning in real estate appraisal. The tool is built with the Django web framework, and it aims to provide researchers with a database-agnostic solution to utilize various machine-learning models in the appraisal. The platform can generate a PDF appraisal report containing the predicted price of the input property and details about the appraisal, including five comparable properties which are automatically extracted from a large database based on relevance. The tool uses the Random Forest (RF) model by default with the ability to swap models if required. Part of the research also focused on evaluating various machine learning models for price prediction using the Ames housing dataset. This involved comparing the accuracy and R2 scores of Linear Regression, Neural Network, and RF models. It was found that the RF model performed price prediction with the greatest R2 value, therefore it was selected as the default model for the platform. The R2 scores for the price prediction model were consistently above 0.9, indicating good accuracy. Finally, methods for optimization of the model including hyperparameter tuning and the addition of bagging were tested and discussed.
Paper
Full text
REALM: Automating Real Estate Appraisal with Machine Learning Models
Semantic Scholar · Computer Science · 2023
Abstract
This paper presents REALM, an open-source tool for utilizing machine learning in real estate appraisal. The tool is built with the Django web framework, and it aims to provide researchers with a database-agnostic solution to utilize various machine-learning models in the appraisal. The platform can generate a PDF appraisal report containing the predicted price of the input property and details about the appraisal, including five comparable properties which are automatically extracted from a large database based on relevance. The tool uses the Random Forest (RF) model by default with the ability to swap models if required. Part of the research also focused on evaluating various machine learning models for price prediction using the Ames housing dataset. This involved comparing the accuracy and R2 scores of Linear Regression, Neural Network, and RF models. It was found that the RF model performed price prediction with the greatest R2 value, therefore it was selected as the default model for the platform. The R2 scores for the price prediction model were consistently above 0.9, indicating good accuracy. Finally, methods for optimization of the model including hyperparameter tuning and the addition of bagging were tested and discussed.