external image airbnb_logo_detail.png

HW3 Airbnb - Discover the World - Dylan Harris


Introduction

Airbnb is the destination for travelers and businesspeople alike. It offers a service that helps its users feel at home across the world and transforms ordinary people into hosts of their own "pop-up" bed and breakfast. One thing that is missing is the step prior to booking the Airbnb. Where would I like to go? I just planned a month-long vacation to Europe and for the last week or so, I had no idea where to go. My solution? Google. But the problem with that is that you are trusting your travels to a virtually faceless journalist who may value certain things that you put zero weight on. What if Airbnb could seize this opportunity to recommend places and stays to its users? In this proposed experiment, I will explain how and why Airbnb should do this.

PHAME Analysis

Problem

The problem is that there is no recommendation system in Airbnb. Sometimes people who want to take vacations have no idea where they would like to go. For business travel, many people would like a similarly structured stay to reduce the amount of variables in their decision-making process. In short, there is no way for a user to be shown a place in the world or an Airbnb that they might like.

Hypothesis

My hypothesis is that many people do not know what they want, and in this same vein, do not know where they want to go. What is initially a long-desired trip to Hawaii is merely a socially designed inclination for a beach vacation. Try South of France, or the Bahamas instead. Places that the user may not even know they want to go! Thus, if Airbnb were to create a recommendation system for vacation destinations as well as the Airbnb stay, users will use Airbnb more and first-time users will feel even more “at home” – a value of Airbnb.

Action

Using Machine Learning methods such as clustering and multivariate regression among others, we can learn what similar users like and try to deduce what type of stay an arbitrary user, call her Ashley, would enjoy given the preferences of other users similar to her. Ashley may be looking to go on a ski trip but does not know where to go. This is where the recommendation system comes in and queries for users with similar preferences (given by clicks and filter choices as well as prior history) and returns a list curated by data for Ashley.

Now, as seen on the main Airbnb site, most stays can be separated into “Travel” and “Business Travel.” For the former, we can create a sort of “Discover the World” page, where Ashley can plug in a few preferences for weather and preference for the outdoors vs. the city, etc. This allows Ashley to use the data even if she has never booked and Airbnb. Anyways, once Ashley has plugged in this information, we can then use her previous travels and filters–if any–to query our offline machine learning models. Then, recommend a few places with a few Airbnb stays per place.

For “Business Travel,” Ashley already knows where she is going, so let us use the Airbnb data to instead find a stay that she would like. More specifically, using Ashley’s prior filter preferences and similarity to other users, we can deduce a list of stays she may like.

Measure

This is simple. We can test the effectiveness of this experiment using A/B tests and surveys. The A/B tests would give us statistical analysis for metrics associated with user engagement with recommended content. For example, “time spent on Discover the World,” or “number of times recommended place booked” are just a couple of the possible metrics that we can gather. The surveys can be used for possible improvements and error detection (“I wanted to ski, why did you recommend St. Tropez?!”).

Evaluation

The evaluation will be a combination of both of these possible sources of truth with a heavier weight toward the results of the A/B tests.


New Data

Sources

The data should already exist in the Airbnb database. The work will be using this data to train models for our recommendation system.

Feature Extraction

Features should be split up into two categories: past preferences and current preferences. Past preferences will be a will be a vector of binary metrics for destinations visited and filters chosen. Current preferences will be a vector of the preferences users choose at the "Discover the World" page. This will narrow down the search space to places that satisfy these requirements while allowing the past preferences to curate the places.


How this Feature will Improve Airbnb

Enhance Engagement

I believe this feature will make users feel more engaged and welcome on the site. It will make it easier to navigate because this feature eliminates the need to pick a destination before coming to Airbnb. More engagement = happier users.

More Bookings

This is the Amazon effect. If Airbnb knows its users better based on data analysis, they will have a better insight into what its users want and which bookings they are more likely to pick. If Airbnb shows these listing first, probability holds that there will be a much higher expected revenue.


Airbnb - Discover the World


Introduction

Airbnb is the destination for travelers and businesspeople alike. It offers a service that helps its users feel at home across the world and transforms ordinary people into hosts of their own "pop-up" bed and breakfast. One thing that is missing is the step prior to booking the Airbnb. Where would I like to go? I just planned a month-long vacation to Europe and for the last week or so, I had no idea where to go. My solution? Google. But the problem with that is that you are trusting your travels to a virtually faceless journalist who may value certain things that you put zero weight on. What if Airbnb could seize this opportunity to recommend places and stays to its users? In this proposed experiment, I will explain how and why Airbnb should do this.

PHAME Analysis

Problem

Hypothesis

Action

Measure

Evaluation


New Data


How this Feature will Improve Airbnb