The Impact of Dimensionality Reduction Techniques on Real Estate Appraisal Performance in Tree-Based Machine Learning Models

Abstract

Accurate and scalable real estate appraisal increasingly relies on machine learning models trained on rich, high-dimensional tabular data. While dimensionality reduction (DR) is often recommended to mitigate noise and multicollinearity, evidence on when DR actually helps tree-based estimators remains mixed. This study provides a systematic, model-level assessment of DR for automated valuation using structured residential attributes (physical, locational, neighborhood, temporal). We benchmark three widely used tree-based learners-Decision Tree, Random Forest, and Gradient Boosting-under three input representations: (i) the full feature set, (ii) supervised feature selection using Random-Forest importance (top-k) and (iii) unsupervised projection via principal component analysis (PCA). Performance is evaluated on held-out test data using coefficient of determination (R²) and root-mean-square error (RMSE). Results indicate that ensembles (Random Forest, Gradient Boosting) already handle moderate dimensionality well, so aggressive feature culling can slightly erode accuracy; by contrast, a single Decision Tree benefits marginally from a compact, high-signal subset. PCA consistently reduces accuracy relative to the full feature set for all tree models, reflecting the fact that the highest-variance directions in features do not necessarily align with price-predictive directions. A practical implication for mass appraisal is that dimensionality reduction should be “task-aware”: embedded/selection methods tied to the target can be helpful when models are capacity-limited, whereas unsupervised projections risk discarding valuation-relevant information. We close with guidance on when to prefer full features, selective pruning, or learned representations in property valuation pipelines.