Science Stories

Discover science successes with XSEDE. Read feature stories where XSEDE has helped researchers in their work.

Key Points
See how XSEDE supports science
Contact Information
« Back

Image Feature Extraction and Demand Estimation on Airbnb: A Deep Learning Approach

Image Feature Extraction and Demand Estimation on Airbnb: A Deep Learning Approach

Shunyuan Zhang, David A. Tepper School of Business, Carnegie Mellon University

The "global sharing economy"—in which nonprofessionals offer paid services via the Web—could add as much as $335 billion to the world's economy by 2025. But because its hosts are not hospitality professionals, buyers face more uncertainty over the quality of a given lodging than with a known hotel chain. It seems intuitive that a welcoming photograph of a room for rent can help overcome that uncertainty, and that factors identified by research on art and professional photography, consumer behavior and psychology can offer guidance on what distinguishes photographs that will or will not improve sales. But can we stipulate a more fine-grained set of rules, and could a machine-learning approach help identify and automate production of more effective photographs? Working with advisor Professor Param Vir Singh and colleagues at the David A. Tepper School of Business at Carnegie Mellon University, graduate student Shunyuan Zhang created a "deep learning" program to analyze 22,000 real-estate-sales room images picked at random from a home-renting website for factors that associate with better sales. Using the GPU and CPU nodes of the heterogeneous, XSEDE-allocated Bridges system at the Pittsburgh Supercomputing Center, the program analyzed the data pixel by pixel, creating a set of rules for effective photographs. Importantly, the machine-learning system identified the same factors previously identified by research on art and professional photography, consumer behavior and psychology. But it identified many more in addition, as well as returning very fine-grained measurements of how these affect sales. In all, the program identified 470 million parameters that associate with demand for a property, with the best photographs associated with a 17.5% greater demand than the worst. Some of these parameters require compositional elements necessarily addressed at the time of photography; others, however, may be incorporated into an automated tool for post-photography processing of the image. Zhang reports her results in a paper now under submission at a peer-reviewed journal.


Figure 1: Photos of several bedrooms, before (left) and after (right) applying the machine-learning algorithm's recommendations for improvement.