Recommender Systems - Airbnb
Scenario Design
Recap from Discussion Week 10
Perform a Scenario Design analysis as described below. Consider whether it makes sense for your selected recommender system to perform scenario design twice, once for the organization (e.g. Amazon.com) and once for the organization’s customers.
Target Users
The target users of this site are people (potential “guests”) who would like to rent rooms or houses on a temporary basis, for work, for pleasure, for a vacation, or for other reasons. Since no rentals can take place without landlords or owners (“hosts”), Airbnb must cater to this type of user as well.
Key Goals
The key goals of the site are twofold:
A. to provide a place for potential guests to peruse accommodations, and
B. to provide a place for hosts to post their property.
Simply put, the goal is for guests to book accommodations outside of more formal hotels and resorts.
Methods
The goals are accomplished through implementing an intuitive design. The site begins with a prompt that requires the entry of the most important information:
A. WHERE
B. CHECK-IN
C. CHECK-OUT
D. GUESTS
With the limiting factors set, the user–obviously a potential guest, not a host–immediately begins shopping through listings of available accommodations. I searched for accommodations in Lisbon with a check-in date of June 12th:
From the perspective of the host, posting available accommodations are every bit as intuitive. By ascertaining the important, limiting information at the outset, one can avoid overwhelming the user with too much information and provide a comfortable amount of data in each page. The importance of matching search results is obvious when one considers the multitude of listings possible for each search.
Reverse Engineering
Attempt to reverse engineer what you can about the site, from the site interface and any available information that you can find on the Internet or elsewhere.
To accommodate the needs of two types of users–leasers and leasees, landlords and tenants, hosts and guests–various methods exist to recommend listings.
In the article “Listing Embeddings in Search Ranking”, Mihajlo Grbovic, et al. describes the strategy behind a recent improvements made to the recommender system for guests. By implementing a “Listing Embedding Technique”, user interactions with the site are retained and patterns recognized as if they were single word tokens within a context. Click and page view sequences, as well as key words associated with those clicks and views are encoded into vectors and sequences of vectors. As interaction increases–within limits set to optimize speed and efficient use of resources–the association between successive clicks is established and “next” page views are recommended based upon the resulting model.
Based on my work in Project 4, I visualize this model as an analysis of page click frequencies (instead of word frequencies in Project 4) paired with time spent on each page; the longer the time spent on a page, the more “important” the attributes associated with the listing and the more compatible the guest is with the host. Additionally, the trajectory of the interaction with pages through each session can be modeled, itself a pattern open to a predicted destination or final page view (a “booking”).
Rather than simply modeling on preferences at one point in time, the sequence of clicks suggests a direction and trend toward certain clusters of characteristics within a session; the algorithm observes selections within a time frame and recommends accordingly, all within the parameters set on the Airbnb home page.
Additionally, an algorithm should predict which renters are most likely to be acceptable to hosts based on several available data points; in the article “How Airbnb uses Machine Learning to Detect Host Preferences”, Bar Ifrach explains his motivation behind an independent research project examining host preferences: “We set out to associate hosts’ prior acceptance and decline decisions by the following characteristics of the trip: check-in date, check-out date and number of guests.” The author wished to develop a more effective recommender system for guests with the addition of host preferences to a model already formed solely on guest preferences; the innovation would combine preferences of both. The host data proved difficult to adapt to traditional methods with data on hosts comparatively sparse and somewhat limited to a simple binary acceptance status.
Additionally, native tools within the application limit guest-host pairing; hosts may set parameters on guest characteristics they are willing to accept.
In the brief article “Airbnb Content-Based Recommendation System”, a rudimentary recommender system is programmed based on Airbnb data. Although limited to accommodation searches in Seattle, the code provides a glimpse into the mechanics associated with recommender systems. The model is built upon term frequency - inverse document frequency and eventually cosine similarity.
Recommendations
Include specific recommendations about how to improve the site’s recommendation capabilities going forward.
Parameters for guests and hosts should be expanded; with more specific parameters set, recommendations could become more precise. More data collected and displayed would only strengthen the efficiency of the models developed on user interactions. If frequency of Airbnb use is not considered, it should be; those guests who are frequent Airbnb renters should be sent to the pickiest or most reticent hosts.
Additionally, algorithms for hosts should be developed for heavily populated cities, such as NYC, LA and Chicago. With surges in tourism during particular times of year, a hierarchy for hosts would be helpful to match hosts with guests they are most likely to accept.
Interest in travel also spreads to location-specific data that can be captured in various ways; rather than depending on the host to list attractions or landmarks within walking distance, specific geolocation data can be integrated into the search, assisting with recommendations and informing general marketing strategies. Guests with several annual trips “for pleasure” or with specific entertainment interests could be targeted for offers or appropriate ad content.
Resources
“Listing Embeddings in Search Ranking”, Mihajlo Grbovic, Haibin Cheng, Qing Zhang, Lynn Yang, Phillippe Siclait and Matt Jones
“Airbnb details its journey to AI-powered search”, Kyle Wiggers
https://venturebeat.com/2018/10/24/airbnb-details-its-journey-to-ai-powered-search/
“How Airbnb uses Machine Learning to Detect Host Preferences”, Bar Ifrach
“Airbnb Content-Based Recommendation System”
https://www.kaggle.com/rdaldian/airbnb-content-based-recommendation-system