using SAS Enterprise Miner 🖥⛏
Airbnb, a platform for short-term lodging, faces challenges such as regulatory scrutiny and criticism for its impact on housing prices and community life. The key issue is distinguishing between hosts offering legal short-term rentals and those running illegal, hotel-like operations. To address this, a comprehensive analysis using cluster and classification techniques on Edinburgh's listing data will be conducted. The goal is to identify patterns that differentiate genuine short-term rentals from businesses, contributing valuable insights to community decision-making and addressing urgent housing concerns. This project was carried out in SAS Enterprise Miner.
Skills and tools used
- Exploratory Data Analysis (Variable Identification, Univariate Analysis, Bi-variate analysis)
- Correlation Analysis
- Variable Reduction (Principal component analysis)
- Cluster Analysis
- ML modelling (decision tree)
- Model Performance evaluation
- Context
- Objectives
- Data Description
- Process Flow
- Initial Data Exploration
- Data Pre-processing
- EDA
- Dimensionality Reduction
- Model Implementation
- Analysis
- References
Airbnb facilitates short-term lodging and tourism-related activities, allowing guests to search for accommodations using filters such as location, price, and property type. Users provide personal and payment information before booking, and some hosts may request government-issued identification. Hosts provide details about their listings, including the number of guests, property type, amenities, and minimum nights for a reservation. The platform incorporates a review system where hosts and guests can share their experiences and rate each other. However, concerns about future stays may impact the truthfulness and impartiality of reviews. Airbnb's policy requires users to forego anonymity, potentially affecting users' willingness to leave negative reviews. Criticism of Airbnb includes its impact on housing prices, nuisances, and security issues in cities where it operates, leading to regulatory scrutiny in various regions, including San Francisco, New York City, and the European Union. The sharing economy, of which Airbnb is a part, presents both opportunities and challenges for homeowners, communities, and governments. One significant issue revolves around whether hosts are operating permanent hotel-like rentals (illegal short-term rentals) or occasionally sharing their primary residence (legal short-term rentals). The responsibility to address housing shortages, nuisances, and safety concerns falls on communities, necessitating a detailed analysis of listing data from Edinburgh to differentiate between legal and illegal short-term rentals. To achieve this, a cluster analysis (unsupervised learning) will be conducted using detailed listing data (listings.csv). This analysis aims to identify meaningful patterns and rules based on host information, property characteristics, and overall guest ratings. Additionally, a classification model will be developed to differentiate between genuine short-term rentals and those operating as businesses. A comprehensive literature search will complement data exploration and preparation for the clustering and classification tasks.
The goal of this analysis is to identify patterns proving/disproving a host’s involvement in illegal rental activities. The domain knowledge gained on the sharing economy (the business domain to which Airbnb belongs) will be used throughout the report in making decisions on variable classification, cluster interpretation and model evaluation. A key takeaway from the project will be to understand how the model outcome can be integrated with Edinburgh regulations for possible implementation. Guest characteristics, Airbnb’s limitations and host activity will be used to gain insight on activity. Ultimately, this project will contribute to the efforts of platforms like Inside Airbnb in improving the sharing economy in Edinburgh. Some questions of interest which will be answered after modelling are:
- Are listings with high occupancy rate likely to be illegal?
- Are certain property types more or less likely to be illegal?
- Are verified hosts more or less likely to engage in illegal activity?
Airbnb does not release data on its listings however, InsideAirbnb scrapes publicly available data about listings from their website. The Edinburgh data (16 December 2022) used in this project is available to download from InsideAirbnb. The listings data has majorly been used for modelling while the calendar and reviews data was used to gain understanding of the measurement parameters.
- Missing values are detected in several character and numerical columns
- Several numerical variables had skewed distributions.
Data cleaning was done following defined criteria in Microsoft Excel using filters. There were 7,387 observations with 25 columns post cleaning.
- Numerical variables accommodates and price have a minimum of 0. It is highly unlikely that a listing on Airbnb would be free and accommodate no-one, so this was regarded as poor data. Therefore, the two corresponding listings were dropped.
- reviews_per_month and review_score_rating had a large number of missing variables, this was permissible because not every guest leaves reviews.
Resulting Summary statistics of Interval and Class Variables for analysis:
The filter node was applied to have an understanding of the outliers that exist. No class variables were excluded at the original 1% cutoff. However, the interval variables were heavily filtered based on their deviation from the mean.
- This is undesirable for this business case as the model needs to be trained to have the ability to accurately identify illegal listings based on specified criteria as it sometimes may be presented with listings that have extreme values.
- For instance: post-filtering only ≤ 4-bedroom listings were left. This excludes several property types (condo, bungalow, rental unit, entire home) from the analysis.
- The results from this node were therefore disregarded.
- All class variables (except binary) transformed using Group Rare Levels @ 4% cutoff.
- Max.Normal applied to review_scores_rating and longitude
- Log10 transformation on all other numerical variables
- No transformation on availability_365 and latitude
To improve illegal activity predictions, occupancy_rate and listing_income have been introduced as new derived variables.
occupancy_rate - days booked divided by the days available to be booked (Airbnb, 2023)
(reviews_per_month * minimum_nights * 12)/availability_365
listing_income - occupancy rate multiplied by the nightly price (An, 2023)
price * occupancy_rate * 365
These variables were log transformed for normalization as their formulas were derived from initial variables that produced skewed results
Price Frequency
Property Type Frequency
Neighbourhood Location
Listing income frequency
Occupancy rate frequency
- still experiencing slight skewness and kurtosis which can be attributed to the skewed distribution of minimum_nights_avg_ntm in its formula.
- strong positive correlation between bedrooms and accommodates. Indicating that they are very similar variables.
- bedrooms dropped to prevent model overfitting.
All non-numerical variables (to prevent abnormal results) were rejected including bedrooms (based on the correlation results).
- 69% similarity between the original 13 variables and the new 6 principal components. This similarity is accepted being that the goal is variable reduction and not similarity or correlation
Standardized and range clusters were formed using Average, Centroid and Ward measurement methods to test for best application. 90% sampling was done to test stability and it produced the same number of clusters on the full and sampled data.
- Clustering analysis focused on listings with similar traits, using new principal component variables and applying range and standardization techniques to prevent abnormal distributions.
- Clusters with less than 1000 observations were disregarded to ensure meaningful results, leading to the rejection of initial clusters with large frequency spreads.
- Despite challenges with standardized clusters, stable results were obtained by specifying two clusters using range measurements, particularly with the Ward method, which demonstrated consistency through sampling.
- The analysis and interpretation of the stable clusters were guided by relevant tables and knowledge of the sharing economy, providing insights into the characteristics of the identified clusters.
- Two clusters formed, each indicating the presence or absence of illegality.
- Target signifies illegal listings
- The meta data node was used to set the target as a binary variable (1= true, 0 = false).
- The dataset was split into 40% train, 30% validation, and 30% test.
Maximum depth = 6
- 96%, 93% and 92% accuracy respectively.
- Snapshot of predictions
Variable Importance: PC2, PC1 and PC3 were the most predictive in model training and had importance values of 0.79 – 1.00
- 91% precision, 94% recall
Model comparison:
Q1: Are listings with high occupancy rate likely to be illegal? The below histogram shows that occupancy rate is not a strong determinant of illegality as listings with low occupancy can either be legal or illegal depending on other factors.
Q2: Are certain property types more or less likely to be illegal? The distribution below shows that any type of properties can be used for illegal rentals. Though a slight distinction can be noticed for ‘entire homes’ showing that they are more likely to be illegal but this is not a definitive factor given that the probability starts from 60%.
Q3: Are verified hosts more or less likely to engage in illegal activity? Similar to the above different hosts can be involved in illegal activity.
The above plots show that a singular variable cannot exhibit traits for illegality as it is influenced by a combination of a series of actions carried out by hosts, which then create a pattern.
Implication of predictive model to Airbnb The number of false positives and negatives on the training and validation set is low. The lower the false negative, the better or else it would defeat the business case/data mining goal.
Based on the validation data:
- 3% false negative = 68 illegal listings falsely classified as legal
- 4% false positive = 102 legal listings falsely classified as illegal
Recommendation for mandatory licensing on the platform Edinburgh laws introduced a license for short-term rentals. However, from analysis of the data, only 0.2% of listings had obtained and declared their license on the platform, though the law is meant to be fully enforced by July 2024 (Edinburgh.gov.uk, 2023). After this date, if Airbnb makes it mandatory to obtain this license for operation, in addition to an activity detection tool (like this classification model) on their website, it is very likely that the illegal listings (including the false negatives) will cease to exist in Edinburgh. This is because hotels/commercial properties do not qualify for the short-term license. Hence, they would be unable to prove to Airbnb that they are short-term rental hosts. Meanwhile, the false negative and false positive listings with a license may be subjected to in-depth manual individual evaluation (if Airbnb has the labour) to prove their validity. Alternatively, these listings each make up 3 and 4% of the population, so the business may decide to disregard them. Meaning that false positive hosts may lose their source of income and impact Airbnb financially, while false negative hosts will continue to operate. Therefore, the impact on the rental market will greatly be reduced to only ≈ 3%. Despite this, some hosts may be able to avoid detection by creating multiple guest accounts and illegally renting their property.
- Airbnb (2023) How do I read my performance data for occupancy and rates? - airbnb help centre, Airbnb.
- An, J. (2023) Airbnb calculator: Host profit estimator: Airbtics: Airbnb analytics, Airbtics.
- Edinburgh.gov.uk (2023) Licences and permits applications, The City of Edinburgh Council.