Zillow.com Recommender Systems

Scenario Design Analysis

• Target Audience: Individuals looks to purchase or rent real estate properties. Types of properties include apartments, houses, townhouses and lots.
• Key Goals: The key goals are to match an end user with a property that closely fits their needs.
• Means to Accomplish Goals: The use of recommender systems to provide the end user with properties they have not seen that meet their search criteria. This tool is especially useful when combining with email notifications. Emails allow for zillow to send new properties that are listed that closely match previous searches.

Reverse Engineering - Recommender Techniques Used

The blogpost “Home Embeddings for Similar Home Recommendations” by Sagdi Lin provides a break down of the techniques use to provide recommendations to its end users. Zillow makes use of of categorical data and numerical data in their recommender systems. Categorical data sets are zip code, school district, home type, zoning type and location. Numerical attributes are price, sqft, lot size, number of bedrooms, number of bathrooms and year built. Once a baseline search is performed neural networks are used to provide find similar properties.

Reference Blog: https://www.zillow.com/tech/embedding-similar-home-recommendation/

Recommendations to Improve Recommender Systems

From a personal point of view, when using zillow to search for a new apartment to rent, I found the the recommendations to be fairly close to what I was looking for. The emails sent are a great idea because it encouraged me to go back to the website to find other apartments. In doing searches for sales of homes, one area that I did find could have improvement is the condition of the home. Searches in the New York region yield a lot of the properties are 30+ years old. Depending on the goals of the end user (buy/flip vs permanent residence) it would be beneficial to categorize recent renovations. This can be done implicitly by lowering or increase the value of the home you are searching for, but having this built into the recommendations would be an improvement.