The reason the talks broke down between Zillow and our local Sandicor MLS is because Zillow wants to keep and own the data we provide. The board members turned them down, but Zillow has made deals with most brokerages independently, which probably include them keeping the data.
You can see where they are going – they want to use their algorithms to predict everything. Eventually, won’t they be telling you how much to pay?
Data scientists at Zillow Group are developing complex computer programs that detect specific attributes in photographs of homes, which could aid in estimating their value.
Advances in deep learning, big data and cloud computing have converged to allow the online real estate database firm and others to develop technology that mimics how the human brain processes visual images–a concept still in its early stages and once limited to only the largest technology companies.
It’s a daunting task, but one that could have wide-ranging impacts as enterprises look to extract meaning from giant databases of photographs.
Digital photographs from sources like real estate listings and online photos are invaluable in accurately estimating a home’s value, Zillow Group executives say. Humans are far better than computers at recognizing the features within a photo that could boost price, but the challenge is sifting through those photos at scale.
Data scientists at the company are developing convolutional neural networks, computer systems designed to mimic the human brain, trained to correlate specific collections of image pixels with valuation signals. For example, if granite countertops and stainless steel are identified in a photograph, that will automatically signal an increase in price.
“It’s essentially trying to code, in the pattern of a neural network, what we as humans can infer just by looking at the image,” said Stan Humphries, chief analytics officer, who heads the firm’s 100-person analytics division.
Neural networks require heavy computing power and in order to develop the technology, Zillow Group uses graphical processing units and cloud-based services from Amazon Web Services. The affordability of cloud-based services for enterprises has made it possible to develop technologies like neural networks that were previously unattainable.
“There were specific algorithms that we prototyped in 2005 that we rejected because it was computationally infeasible,” Mr. Humphries said. “Today, I can’t remember the last time I rejected a prototype on the basis that it was too compute-intensive.”
The deep-learning approach on image data could be integrated into the company’s valuation algorithm, called a “Zestimate,” during the first quarter of 2017, he said.
The Seattle-based firm has amassed a database of 115 million homes across the country. Zestimates are used to estimate each property’s valuation, based on statistical and machine learning models that examine hundreds of data points on each home, including square footage, lot size, number of transactions in a geographical area, and soon, hundreds of thousands of photos. Since 2005, the company has reduced its valuation error rate from 14% to 4.5% through iterations of its algorithm, and it’s betting that estimates could be even more accurate with sophisticated neural networks.