In the field of real estate, the idea of predicting the “right price” for a property is growing heavily in interest. Most current algorithms solely use statistical information about given properties as a form of input to predict its right price. However, these algorithms fail to include a notable form of data that often influences the perception of a buyer: images of the house. Fortunately, in recency, convolutional neural networks (CNNs) have increased in prominence for their ability to generate strong feature representations out of images and use those representations to accurately map to scalar/vectorized outputs.
Yet, as explored in the 2021 project, creating a dataset of strongly-labelled pictures and corresponding statistical information for homes in the Greater Toronto Area (GTA) w.r.t. houses sold at a single point in time is a difficult task. A MAPE error of 15% was generated through extensive hyperparameter tuning on a dataset of 500 houses. However, this number is far behind current MAPE benchmarks developed through numerical datasets only. Furthermore, on the numerical side of things, high quality GTA residential datasets that contain a wide variety of house attributes are far and few to be found.
The lack of high-quality GTA datasets lead to the core goal of this sequel RealValue project, which is to create a dataset centered around GTA homes that is extensive, usable and informative. The metric to measure success in executing this goal is the MAPE that different experiments are able to generate.
The project is considered a success if the project achieves 7.5% MAPE error, which is 1/2 of the best error from the 2021 project.
If this project and its goal interests you, please reach out at email@example.com.