Airbnb Deep Learning Project

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

What is a good image for an Airbnb property? We analyze this question. Airbnb is a sharing economy platform where individuals can list or rent lodgings. Lodging is an experience product, Airbnb lists a wide variety of properties, the quality varies immensely across properties, hosts are not professionals, and as a result guests face huge quality uncertainty. In this scenario, property images play an important role in reducing quality uncertainty and improving property demand. While images are central to the success of Airbnb, hosts have little idea on what makes a good image for an Airbnb property.  Using a unique and detailed dataset from Airbnb, we propose and estimate a deep learning framework where demand is modeled as a function of images, reviews, price, and property as well as fixed effects. Parallel Convolutional Neural Networks (ConvNets) are embedded to extract features from property images. This framework incorporates an econometric model while making it feasible to process images, which represent high dimensional and unstructured data. The results from model comparison shows that incorporating images improves the model’s prediction power by approximately 33%. With the unique architecture of our deep learning model, we are able to identify differential impact of images. The results indicate that bedroom images are valued the most followed by bathroom, living room and kitchen. Within each room type, there is a huge variation with the best bedroom pictures providing 17.5% greater demand than the worst bedroom pictures per year to a property. We rank room images based on their impact on property demand, and summarize the differences between high ranked versus low ranked images. We then show a process to interpret which image features are captured by a ConvNet filter. We show how to optimize a property photo given our framework. We focus on image features that can be automatically optimized during photo post processing. We report how much a feature could impact the demand of the property for several features including image composition features such as diagonal dominance and rule of third, and post processing features such as illumination, color tone, and color curves. Overall, results indicate that features that improve illumination of the image, or make it look more spacious or open, or provide a sense of warmth or orderliness/purity have large impact on the demand for a property.