Realtor NN

PCA (8 components, 84% of variance) and feedforward neural networks classifying U.S. residential properties into price classes. Low and Medium are easy; High and Very High are the hard ones.

Overview

I apply Principal Component Analysis (PCA) and feedforward neural networks to classify U.S. residential properties into price classes, using a large real estate dataset.

Goals

• Reduce the dimensionality of the original feature space while controlling reconstruction error.
• Train neural networks for multi-class price classification on the reconstructed feature matrix.
• Compare several architectures and training setups using mean squared error (MSE) and per-class performance metrics.
• Implement creative extensions beyond the baseline workflow and discuss their impact.

Results

• PCA with 8 components captures about 84.35% of the variance, with a reconstruction MSE of about 0.1565 in the standardized space.
• The data is reconstructed back to the original standardized coordinates, so the networks take 11 inputs and each input neuron stays an interpretable reconstructed feature.
• Of the architectures tested, the deeper 11-32-16-8-4 network reached the lowest final MSE (about 1.73), a modest but consistent improvement over shallower and smaller models.
• Performance is uneven across price classes: Low and Medium are relatively easy to classify, while High and Very High remain challenging.

The repository contains the project notebook.