This is part of the course project for MIT 15.095 Machine Learning via a Modern Optimization Lens.
The objective of this project is to analyze the available data for the following:
- Characterize the different kinds of properties/listings in the data
- Predict price of a listing
From Kaggle (www.kaggle.com/datasets/whenamancodes/london-uk-airbnb-open-data), we have the following data available for ~70k AirBnB listings in London UK:
- Property Identifiers: Property ID (numeric), property name, neighborhood, lat-long coordinates, licensed or not
- Host Details: Host ID (numeric), host name, number of listings the host has in the database
- Listing details: Room types, availability information, price, minimum nights for a
Price prediction at overall level:
- We use stratified sampling to retrieve 20% of the data as testing data, ensuring good representation of neighborhood categories and room types in test data
- We use stable regression to split the remaining 80% data into training data (60%) and validation data (20%)
Clustering:
- We cluster the entire data set
- We extract data for each cluster and divide the data randomly into training (80%) and test (20%) sets
Optimal Classification Trees method is used to interpret clusters.
| Cluster | Feature |
|---|---|
| Cluster 1 | Mid-low neighborhoods, offering private or shared rooms |
| Cluster 2 | Listings in very expensive neighborhood (Camden, Westminster, Kensington and Chelsea), that offer either private or shared rooms; AND listings that offer entire homes or are hotel rooms in expensive/middle/low cost neighborhoods but are booked for the most part of the year (available <42.5 days in a year) |
| Cluster 3 | All other properties |
| Cluster 4 | Listings offering entire homes or are hotel rooms that are generally available (availability >201 days) |
| Cluster 5 | Listings with private/shared rooms in mid-high/middle /low cost neighborhoods |
| Cluster 6 | Listings that offer entire homes or are hotel rooms in mid-high neighborhoods and but are booked for most part of the year (<42.5 days in last year), or are in other neighborhoods but have very high availability (between 42 to 201 days in a year) |
- Hosts can use the OCT to understand what makes a property similar to the one they own. The OCT also helps determine how availability varies between different neighborhoods
- The XGBoost model builtt at the overall level can be used to predict prices for listing based on their properties, and can inform hosts when setting prices. It also indicates guest willingness to pay based on the different features like availability
For more results, please check the 'ML Presentation.pdf'.