Learning with self-attention for rental market spatial dynamics in the Atlanta metropolitan area

The rental housing market plays a critical role in the United States real estate market. Prior studies have used various approaches to model housing rent, such as interpolation, hedonic modeling, and machine learning. However, only a few studies have modeled rental prices based on textual data, which provides rich and contextual information about rental properties. In addition, many models, especially deep learning models, use an end-to-end black box for prediction, which hides the decision process. Such models are difficult to interpret and explain the driving factors. This study builds on our previous work, aiming to develop and evaluate rental market spatial dynamics models combining Long Short-Term Memory (LSTM) networks and self-attention mechanism. We compare the performance of the proposed model with our previous models on predicting rental prices in Atlanta, Georgia, USA. We also use techniques from saliency maps to explain the generated model. Results show that the self-attention-based model outperforms our previous models. The saliency map techniques reveal how the model attends to a different part of the textual information. The predicted results reflect the spatial variation of textual information. Such a model offers practical pricing references for homeowners and renters, and spatial patterns for urban planners and stakeholders.