AN ATTENTION-BASED REGRESSION METHOD FOR REAL ESTATE VALUATION

Location is often regarded as the most important feature in real estate valuation, and it is also difficult to encode it as a model variable. We present an Attention-Based Regression, a method inspired by graph attention network, that leverages the masked attention mechanism to mimic the sales comparison approach adopted by realtors. Our method provides a data-driven approach to encode property location that does not require specific neighborhood knowl-edge, allowing the model to cover larger geographic regions. We compare our method to benchmark models and find that our method has performance that matches state-of-the-art boosted ensembles.