Problem statement: Can we build a smarter pricing model for Airbnb hosts to increase their revenue? Can we recommend similar successful listings?
I focused on popular Airbnb listings in Los Angeles, CA with high occupancy rates(>65%) and deployed the project via a flask web app on Heroku. With tens of thousands of listings I determined which factors were most predictive of their price. Using the Inside Airbnb dataset with information on every Airbnb listing and their reviews, I considered many features including the following:
- Bedrooms
- Bathrooms
- Security Deposit
- Minimum Nights
- Review Score
- Availability
- Location
- Property Type
I used linear regression to predict the base price for a given listing. Given the importance of date in setting a nightly price, I used Facebook Prophet to forecast price fluctuations as a function of the date, considering day of week, time of year, and holidays.
To allow hosts, particularly new ones, to compare their listings to similar ones, I built a recommender system using unsupervised learning methods to compare their inputted listing to all other popular listings (those with high occupancy rates).
p05_Data_Clean_Feature_Setup.ipynb shows the process to clean all of the data and prepare relevant features for modeling
p05_EDA_and_Regression.ipynb shows data analysis and the process of training various regression models and evaluating feature relationships and model performance
p05_Time_Series.ipynb shows Time Series Analysis used to forecast price fluctuations as a function of time
p05_Similar_Listing_Recommender.ipynb uses unsupervised learning to build a recommender system to find similar successful listings
p05_NLP_Natural_Language_Processing.ipynb shows analysis of Airbnb listing descriptions, using topic modeling to generate additional features for the linear regression models
p05_Clustering.ipynb is unfinished, but uses clustering algorithms to try and create meaningful, distinct groups of Airbnb listings to improve the quality of the price suggestions
Web App files used to build the browser-based predictor tool hosted on Heroku here
Slides can be found here