Modeling Rent in Atlanta Using Deep Learning

🔍 Why Predicting Rent Matters More Than You Think

Have you ever wondered why rent prices vary so much from one neighborhood to another? Or how landlords decide what to charge for an apartment? These questions aren’t just important for renters and property owners—they’re also key to understanding urban development, gentrification, and even social inequality. That’s why we set out to build a better way to predict rental prices, using cutting-edge deep learning techniques and data from Craigslist listings in Atlanta.

Rent isn’t just about square footage or the number of bedrooms. It’s influenced by a complex mix of factors, from the proximity to public transit to the quality of the neighborhood. Traditional methods for predicting rent often fall short because they can’t capture these nuances. But by combining textual descriptions of properties with spatial data, we’ve developed a model that’s more accurate and insightful than ever before.

🛠️ How We Did It: Mining Craigslist for Hidden Insights

To tackle this challenge, we turned to Craigslist, a treasure trove of user-generated rental listings. We collected data from Atlanta, a city with a rapidly changing rental market, and cleaned it up to remove duplicates and outliers. Each listing included details like location, square footage, number of bedrooms, and—most importantly—textual descriptions of the property.

We then designed a series of experiments to test different models:

• Experiment I: We started with traditional methods like inverse distance weighting and kriging, which use spatial data to predict rent. These methods are great for understanding how location affects price, but they don’t account for the rich details in property descriptions.
• Experiment II: Next, we focused on the text. Using deep learning models like recurrent neural networks (RNNs) and convolutional neural networks (CNNs), we analyzed the short and long descriptions of each property. These models are particularly good at capturing the nuances of language, like whether a listing mentions “hardwood floors” or “close to downtown.”
• Experiment III: Finally, we combined both spatial and textual data to create an integrated model. This approach allowed us to predict rent prices with even greater accuracy.

📊 What We Found: Text Matters—A Lot

Our results were striking. When we used textual information alone, our deep learning models achieved an average error of $196.80 (measured as mean absolute error, or MAE). But when we combined text with spatial and property attributes, the error dropped to just $145.40. That’s a significant improvement, and it shows just how much value there is in the words landlords use to describe their properties.

We also found that certain words and phrases had a big impact on rent. For example, listings that mentioned “modern kitchen” or “walkable neighborhood” tended to command higher prices, while those with vague or generic descriptions often fell short.

💡 What This Means for the Future

Our research has some exciting implications. For renters, it means better tools to find fair prices and avoid overpaying. For landlords, it offers a way to set competitive rents that reflect the true value of their properties. And for urban planners, it provides a new lens for understanding how cities evolve over time.

Looking ahead, we’re excited to expand this work to other cities and explore how factors like gentrification and urban development influence rent. By combining deep learning with big data, we believe we can unlock new insights into the dynamics of rental markets—and help create more equitable cities for everyone.