Application of unstructured text based features in prediction of real estate prices: A comparative study

Abstract

This study demonstrates the potential of application of unstructured textual data for predicting real estate prices and compares different protocols for extracting features from textual data. Performance of the different models for price prediction was evaluated on data set of real estate listings, which included numerical and categorical features, as well as text descriptions. The experiments showed that adding features extracted from both the translated description text, as well as noun chunks from it, resulted in the highest R2 score of 0.768, representing an improvement over the R2 score of 0.71 for the baseline model without text-based features. The findings from this study indicate how the performance of real estate price prediction models can be improved by utilizing text-based features, in turn benefiting property market stakeholders in making informed decisions and evaluating competitive pricing strategies.