Since Seattle-based Zillow first began in 2006, the company has used a proprietary algorithm and team of statisticians to create its famous “Zestimates” for real estate. The company deployed its first version in 2006, and then followed with two major upgrades in 2008 and 2011. It has continued to improve its system with improvements deployed along the way in the form of incremental iterations. The statistician designed software uses an automated process, and has been dependent upon the amount of data available for the location.
The median error rate for the company has been 4.5%, which basically means that half of the homes will be within 4.5% and half of the homes will be off by 4.5% in terms of the accuracy for the selling price. The company uses public records to gather their data, so homes with current information are more likely to have real estate values estimated according to current data trends. The method is a non-linear neural network learning model including multiple variables from data of recent housing unit sales within a given county.
Zillow uses statistics and market data, referred to as data sets, to estimate the market value of homes. Their valuations are automatically computed three times each week, basing the results on the millions of data points submitted by users and collected from public records, such as County Recorders offices and Department of Assessments offices. Real estate valuation is understandably of great interest to the world of financial economics.
The motivation for creating neural network systems is the human brain, and the way it is capable of processing data. The biological neural network is basically an assembly of interconnected units which process information based on the strength of inter-unit connections and either adapting to or learning from sets of training patterns.
Neural Network Basics
To understand what an artificial neural network is, it is possible to use the Microsoft Neural Network algorithm for a brief tutorial. This algorithm has as many as three layers of neurons within a network. These layers include:
- The input layer- defines all input attribute values for the data mining model, and probabilities – This layer approximates the human neuron dendrites, which carry signals into the nucleus
- The output layer- represents predictable attribute values for the data mining model – This layer approximates the human neuron axon, which carries signals out
- An optional hidden layer- inputs are assigned relative weights for importance and relevance – This layer approximates the human neuron synapses which change size in response to learning and provided the basis for the concept of associative memory
The layers work together to function similarly to human neurons, taking in information and making decisions based upon that information. The difference is that human neurons are capable of much more sophisticated decisions based on vast quantities of varied information, with significantly more variables and complex interactions arising from human emotional and spiritual characteristics which machines do not have. In addition, artificial neurons eventually reach a state of no further changes in learning, though they may continue to function. In a nutshell, the neural network will work efficiently with data, but it can only work as well as the quality of data and parameters input by the operator, and it is dependent upon the amount of variability it must handle. Researchers have discovered that forecasting errors can be higher with high data variability.
An Innovative Use of Digital Photographs
What makes the Zillow neural network interesting is that Zillow Group Data scientists are working with the visual aspects of homes to develop computer programs complex enough to estimate their values. The programs are using photographs and complex technology to mimic how visual images are processed by the human brain. The huge photography databases existing online are being sourced using advances in cloud computing and deep learning to create new, evolved neural networks which focus on detecting specific home attributes as aids for estimating values.
The Zillow computer systems are being trained to correlate the differences among image pixels in photographs. Scientists have developed specific collections and assigned valuation signals to them. When the neural network discerns differences in home construction materials in photographs, it will automatically generate a price difference. A variety of home features within photos will be used to extract valuation signals for the neural network.
The scientist will need to create vast amounts of new code for the neural network to differentiate what humans can easily infer by simply glancing at photographs. It is a huge task, and requires that the company develop exceedingly convoluted network components. The Zillow Group uses Amazon Web Services to access the incredible power required to develop this kind of new technology. Cloud-based, and computer intensive, the graphical processing units are making affordable and possible what was previously not.
Zillow’s Chief Analytics Officer, Stan Humphries oversees one hundred people in his division. His analytics scientists prototyped specific algorithms for this type of technology, but rejected them because they were unfeasible at the time. They were too computer intensive. But the company has been able to access advances in deep learning research, using this approach to integrate image data into the famous Zillow Zestimate algorithm. Due to the vast advances in computer technologies and the extensive yet affordable cloud based services, the company expects that it could successfully begin using the newly developed neural network in Q1 of 2017.
Since its inception, Zillow has used its proprietary incremental algorithm iterations to reduce valuation error rates from 14% to 4.5%. Using its huge, 115-million home database, gathered from properties throughout the United States, the company has successfully implemented machine learning models and statistics of hundreds of varied data points to provide real estate valuations. These examination points have included square footage, rooms, number of baths, lot sizes, recent transactions of similar homes in the area and many other specifics popular with home owners and buyers. The plan for Zillow’s future, is to use its newly developed, and extremely complex, neural network to add photographs as well.
Expected Real Estate Benefits
Humans can see differences in materials used in home construction and décor. In person, and while viewing photographs, the human eye can discern if flooring is wood or tile, if shelves are wood or stainless steel, or if countertops are stone or laminate. The quality of materials is readily quantified and stored in the human brain, and perception of value is either increased or decreased. The challenge for Zillow has been to achieve this level of human perception within their computer systems.
Though neural networks have been in existence for decades, and their function has been based upon human biology, they have never been able to attain the complex capabilities for inference and perception which humans take for granted. What is exciting for Zillow, and the real estate market in general, is that the new Artificial Intelligence (AI) in development may provide substantial increases in valuation accuracy and perhaps, increases in sales transactions.
As businesses seek to do better business, AI technologies are increasingly being adopted. The reason is simple. These AI advances are more closely approximating the way humans reason at a rapid pace. In the same way that the human neuron synapses change size in response to learning, the computer neuron layers are being scientifically changed to provide more accurate learning responses. It’s an exciting time for science and business.