Boosting Machine Learning Models with Explainable AI (XAI) - Insights on Airbnb listings

With a typical machine learning model, the traditional correlation of feature importance analysis often has limited value. In a data scientist’s toolkit, are there reliable, systematic, model agnostic methods that measure feature impact accurate to the prediction? The answer is yes.

Here we use a model built on Airbnb data to illustrate:

• Explainable AI (XAI) technologies
• What can XAI do for global and local explanation
• What can XAI do for model enhancement

XAI — a brief overview

As AI gains traction with more applications, Explainable AI (XAI) is an increasingly critical component to explain with clarity and deploy with confidence. XAI technologies are becoming more mature for both machine learning and deep learning. Here are a couple of algorithm neutral methods that are practical to use today:

SHAP

SHAP (SHapley Additive exPlanations) is developed by Scott Lundberg at the University of Washington. SHAP computes Shapley values from game theory, by assuming that each feature value of the instance is a “player” in a game where the prediction is the payout. Then a prediction can be explained by computing the contribution of each feature to the prediction. Note SHAP has these desirable properties:

1. Local accuracy: the sum of the feature attributions is equal to the output of the model we are trying to explain

2. Missingness: features that are already missing have no impact

3. Consistency: changing a model so a feature has a larger impact on the model will never decrease the attribution assigned to that feature.

SHAP supports tree ensemble, deep learning and other models. It can be used for both global and local explanation. Please refer to Scott Lundberg’s SHAP paper.

LIME

Local Interpretable Model-Agnostic Explanations (LIME) is based on the concept of surrogate models. When interpreting a black box model, LIME tests what happens to the predictions with variations of data, and trains local surrogate models with weighted features. Finally, individual predictions for “black box” models can be explained with local, interpretable, surrogate models.

Please refer to LIME paper: “Why should I Trust You”

Airbnb booking rate model

The model used here predicts Airbnb booking rate. It is trained with data for Los Angeles area listings, obtained from insideairbnb.com. For simplicity, I use a subset of features to train an XGBoost model.

For more information on model design, please refer to https://towardsdatascience.com/predicting-market-rank-for-airbnb-listings-59009a886d6

Insight with global explanation

SHAP summary shows top feature contributions. It also shows data point distribution and provides visual indicators of how feature values affect predictions. Here red indicates higher feature value, blue indicates lower feature value. On the x-axis, higher SHAP value to the right corresponds to higher prediction value (more likely listing gets booked), lower SHAP value to the left corresponds to lower prediction value (less likely listing gets booked).

Here are a few insights gained with global feature analysis:

Who are the most successful hosts?

Using Dependence Plots, we can examine the relationship between feature values and predicted outcome. In the first diagram, as the number of listings a host has increases, we see a decreasing trend of SHAP values. In the second diagram, the x-axis shows host listings count, the color shows listings count of “entire home”.

We can probably derive these type of hosts:

• host with single or a few listings — these are individuals and family, their listing are generally attractive likely due to focus and personal care
• host with 15–60 listings — these hosts have the least attractive listings, they are probably small hotel or motel type of properties that rents out rooms?
• host with more than 150 listings — in the second diagram, we can see as the number of host listing increases to above 50, the predicted booking rate increase substantially (reversing earlier trend). Further, those listings are almost all “entire home”. At the top range, those hosts with over 100 “entire home” listings achieve a booking rate of 75% and above which is far superior to anybody else. Are those professionally managed Airbnb property companies?

Higher cleaning fee or higher price?

Given the choice, should a host charge more on nightly price or cleaning fee? The dependence plot shows feature interaction between price and cleaning fee. The red color indicates a higher cleaning fee. Along the x-axis, as price increases, predicted booking rate decreases, which is expected. Further, we see listing with higher cleaning fee (red dots) tends to stay above those with lower cleaning fee (blue dots).

Therefore, a listing with a higher cleaning fee is actually favorable when it comes to predicted booking rate. A host who shifts costs more to the cleaning fee probably wins by encouraging guests to stay longer and makes listing price seems cheaper on the front page.

More reviews or higher review rating?

In the diagram, we see an increase in review rating (x-axis) leads to higher booking rate. An increase in the number of reviews, however, does not correspond to better booking rate (red dots are scattered vertically). A host is much better off getting a few good reviews than having a lot of mediocre reviews.

Insight for Local Explanation

SHAP force plot can be used to explain individual predictions. For example, we can see that there is a base value (bias term) of 0.01249, with features in red pushes that value to the right, and features in blue pushes that value to the left, with a combined output of 0.58. Therefore, the effect of top feature is quantified on the prediction with local accuracy. The particular listing has a number of strong features values (superhost, low price, entire home, recent calendar updates) which makes it favorable for booking.

The listing below has a number of feature values (high price of $390, long minimum stay of 30 days, long time calendar hasn’t been updated) that makes it less likely to be booked.

LIME method can be used to explain individual predictions as well, it quantitively shows the effect of top features (orange is positive, blue is negative).

The first example shows a listing with mostly positive feature values (it is an entire home offered by a superhost, with lots of reviews).

The second example shows a listing with negative feature values (it requires a minimum stay of 30 nights and charges a cleaning fee of $150).

Insight for Model Improvement

By examining the global and local impact of features, we can often reveal unexpected patterns of data and gain new insights. With further analysis, we may find the root cause to be one of the following:

• deficiency with business analysis
• error with data collection
• data processing improvement (impute and scale)
• or, the unusual pattern is a true reflection of new knowledge to learn

The diagram below shows there are mainly two types of listings with high “calendar_updated” value (red dots). One group, on the left most side of x-axis, who has 0 reviews last twelve month, are essentially stale listing with negative SHAP value and therefore low predicted booking rate. The other red dots are scattered in the upper area which indicates they have higher SHAP value and more likely to get booked. Those are listings that are consistently available and require few calendar updates from hosts. This provides a clue for feature engineering with the goal to distinguish stable listing from training data.

Another example shown here is Skater Partial Dependence Plot which shows the interaction of latitude and longitude features, with vertical axis indicating their effect on prediction. Visually, this 3D diagram can be superimposed on a map of the LA area, which clearly shows central and north area being more popular, and south being the least likely to be booked. This insight can not be gained with analysis of individual features.

Making corrections, adjustments, and gaining new knowledge is part of the iterative model lifecycle which should lead to incremental improvements. XAI can uncover hidden clues and provide critical evidence for that.

Github

A simplified version of the model and XAI code can be found here: https://github.com/seanxwang/XAI-airbnb-booking

What is next

Airbnb listings have rich and informative features such as image and text. Incorporating those into models can greatly enhance predictive performance. XAI with deep learning and vision should be both challenging and rewarding as well.

Predicting Market Rank for Airbnb Listings

Motivation

There are several reasons that make Airbnb data challenging and rewarding to work with:

• Unlike Kaggle, where objectives and metrics are defined, open ended problem definition is a critical data science skill - how to identify a valuable business objective and create analytical framework and modeling solutions around it?
• Rich information including structured data, text and images to be assessed and narrated - how to deal with having too much data and missing certain data at the same time?
• Regular data updates from sources such as insideairbnb.com, which provides critical feedback and enables iterative optimization
• Behind data, are the places, people and their diverse culture, to be interpreted and uncovered. Enhancing Airbnb experience can bring enrichment to humanity and makes it a meaningful and fascinating goal for data scientists

Why predict Market Rank instead of Price?

Airbnb listings contain a rich array of information including structured data, text and images. The convenience and updated availability of data from sources such as insideairbnb.com make them intriguing data science projects. Let’s start by building baseline models.

Before designing a machine learning model, the first question to ask is, what is our goal? The goal should also take into consideration the availability and quality of data. A lot of work published focus on price prediction. However, price may fluctuate widely between different dates, which limits a model’s usefulness. Instead, I will focus on ranking a listing relative to its competitors in the market. It is an important design consideration since relative ranking score captures the relative attractiveness of a listing, and makes a model more general to fit with more data and used for more applications.

Here we define market rank as the relative competitive score of a listing in a defined city or neighborhood. How can this prediction model be used? Imagine if you are a host, not only you can see how well you are currently ranked vs the competition, you can also evaluate the effect of altering features (lower price, add amenities…) to improve your ranking. For Airbnb, the ability to predict listing popularity allows effective recommendation for potential guests (“here are top 10 recommended listings in Hollywood for you dates and price range”), or optimization recommendation for hosts (“you will improve your ranking from top 25% to top 10% if you lower your cleaning fee by $20”). Ultimately, the goal is increasing sales and higher customer satisfaction.

Model design and training target

The better a listing does, the higher its market rank should be. How do we define and obtain “market rank” information for supervised training? We are looking for data that tells us how desirable a listing is, or in other words, how fast it gets booked relative to peers. Although we don’t have actual booking information, we have availability and review data that is updated monthly. Let’s look at pros and cons of these two key metrics:

Availability — Availability by itself may be misleading, since a listing can be unavailable either by being booked or host taking off market. By observing availability change from time A to time B, and look at when host last updated calendar, we can more intelligently interpreting change and derive booking activity for most of the listings

Reviews — The increased number of reviews in the target observation window is a good proxy for how well a listing did. There is still a pitfall, those listings with longer minimum date tend to have fewer reviews (relative to the number of days booked)

As illustrated, booking rate and review increase rate can be estimated by:

• for a target time range, looking at when host updated calendar and how availability changes over time, calculates a booking rate
• For the same target time range, looking at how many more reviews are added and calculates a review rate

Booking rate and review rate compliment each other to provide market ranking indicators for individual listings. They are scaled and normalized to be used as training target for regression models. For further diversification, I trained booking rate model and review rate model based on subsets of listing features, while preventing data leak. The resulting booking rate score and review rate score are then aggregated to a combined score, which is then used to generate market ranking in percentile form quantifying a listing’s relative strength in any city, area or neighborhood.

Developing Baseline Models

For the prototype, I used data for airbnb listings in Los Angeles, with 2019–03–06 as time A, and 2019–12–05 as time B, and 2019–12–05 + 30 days as the target booking time window. For simplicity, I used XGBoost with minimal amount of data processing and tuning.

Below shows sample validation curve for review rate model, which has an R2 score of ~0.6. This would serve as a baseline, with potential for improvement by adding features, data, tuning…

Below shows sample learning curve for booking rate model. The gap between training and validation (which indicates variance) shows narrowing trend with increasing data size. As expected, bias is clearly present as seen from error level, which may be improved with features and algorithm.

Below shows SHAP summary plots for the two regression models. The features are listed top down by order of their impact on model outcome. For each feature, its impact is visualized by its reach on the horizontal axis. Red indicates higher feature value, while blue indicates lower. An extension to the right quantifies positive impact (higher the feature value, higher the booking and review rate, therefore more hot the listing is). An extension to the left quantifies the negative impact. Some quick observations:

• As expected, higher price, fewer reviews, longer minimum nights negatively affects booking rate
• Entire home/apt is most preferred by guests
• Simple features such as coffee machine, self check-in can boost a listing
• Feature impact are different for the two models, therefore making them complimentary to each other

SHAP (SHapley Additive exPlanations) is an emerging algorithm agnostic Explainable AI (XAI) method which leverages game theory to measure the impact of features accurate to the prediction. Detailed SHAP analysis is outside the scope here.

Market Rank Illustrated

Here we evaluate the models using a reserved test set of listings in Los Angeles area. After combining scores from booking model and review model, we obtain a city wide ranking score (“market_rank_city”) in percentile form. For example, listing ID 8570847 is at top 98.6% in terms of its market competitiveness, while listing ID 14014479 is at bottom 6.9%.

Check out actual listings following standard url format https://www.airbnb.com/rooms/, to see if scores make sense.

Market Rank can also be calculated for a defined target group, leading to more applications. For example, we can rank listings in a particular neighborhood and within a price range. We can apply any feature value as filters to compare competitiveness of a listing in a subset.

Below is an example of calculating percentile rank (“market_rank_neighborhood”) for listings priced at $150-$250 in Hollywood neighborhood.

From the above list, let’s look at the top listing, which is a new 2 bedroom condo with full kitchen, priced at $174/night: https://www.airbnb.com/rooms/32792761

And bottom listing, which is a thirteen square foot room with a double bed, priced at $199/night: https://www.airbnb.com/rooms/28012404

Model performance on new listings

Note we include historical information such as increasing rate of reviews and booking activities to predict future outcome. This is not a form of data leak. Rather, it is a true reflection of how a potential guest evaluates a listing. Consequently, model is more neutral on new listings because of missing information on many features. This is also a true reflection of reality.

To more accurately differentiate the quality among new listings, it is desirable to develop models that focus on features such as amenities, photos, text, and location. Think KNN built on subset of features that is constant with time.

What is next

I have shared the journey of a data science project utilizing real world data, starting from a meaningful objective, by researching available data and experimenting with model combination, to promising results. The baseline models developed only scratched the surface of what is possible. Thanks to the vast amount and dynamic nature of airbnb data, further improvements may come from more data scrubbing, feature engineering and algorithm tuning. Adding images and text information also makes for exciting exploration with deep learning.