TRIP ADVISOR
IN-DESTINATION
ASSISTANT
A Machine - Learning Based In-destination Recommendation Assistant
Overview
Date: March 2018
​
Team: Helen He, Amanda Mattson, Justin Siris, Borui Yang, Matt Cocuzzo
​
My Role: UX Design Lead
​
Client: Trip Advisor
Highlights
-
In this project you will see a design for a feature building block that could flexibly sits inside of a established, complex system, and can be deployed by its own with minimum architectural change and budget.
-
The goal of this design is to help those who are in need of a quick suggestion of an interesting tourist attraction or delicious restaurant tailored to their specific tastes close by.
​
-
I worked in a innovative, cross-functional team including 4 students in engineering and business. We combined machine-learning, sentiment analysis, feature vectors, and GPS into user experience for solo-travelers, which get us 3.8/4.0 GPA for this class and well recognition from Boston clients - Trip Advisor User Experience Team.
Process
Exploratory Research
Persona and Scenario
System architecture and conceptual model
Pivoting
Design Solutions
Exploratory Research
First, we defined which part of TripAdvisor we wanted to study, and decided to focus on its personal recommendation section. Then we did research and checked some on-line reports to identify who would be the target users that could bring greater amount of profit and ROI (Return on Investment) to the product. The independent traveler becomes our focus.
Independent Traveler
Tom, 30
​
Resources: TripBarometer for Trip Advisor 2016
Persona, Scenario
Independent Traveler
Tom, 30
​
Travel Routine: For Tom, travel means discovering a bunch of uncertainties and taking a whole mental break. Tom just book the hotel and flight ticket ahead of time, then he likes to hang out on the street, take a walk in the unknown city, and discover interesting places and restaurants on the go.
​
" I don't like to plan travel ahead of time. Researching a bunch of hotels and restaurants is wasting my time and I always change my mind in a second. I'm the type of person who don't want to always been settled and know where I'm going next. Life to me is a discovery journey, the destination is not so important."
User Flow
After we conducted preliminary research and understand the target persona. we discussed the possible user flow that could satisfy user's needs based on our observations and assumptions.
1. Users create their profile and select their interest (foodie, beer lover, etc. ) at the beginning of using TripAdvisor.
​
2. Users opt into TripAdvisor In-Destination Recommendation System.
​
3. User set their preference.
​
4. Based on the user’s data, the system will combine with browsing history and use a hybrid filtering to provide prioritized suggestions.
​
5. When the user travels, he can open TripAdvisor, by scanning the building and turn on the GPS, Trip Advisor will tell him the overall rating and whether that matches this user's preference.
Bill on their way and explore the city randomly
Bill find a good restaurants!
Bill open TripAdvisor
This restaurants have reviewed 4 stars, and provide vegan food that Bill wanted
Conceptual Model
Using research and persona, user flows as a guide, from the backend perspective, we have discussed certain engineering approaches to support the design solutions.
The data comes from reviews and information from the forums. All data taken from TripAdvisor.
First, we will filter out some of the candidates based on the user’s interests and browsing history. The interests and browsing history are customized by user. This filtering is content-based filtering on user’s profile and browsing history.
Then the sentiment analysis will be run on the rest of the candidates and we get the sentiment rating for each place or activity. The sentiment analysis is collaborative filtering so it is based on other people’s reviews and information from the forums. The high sentiment rating captured in user's text would trigger relevant information been recommend to user.
The user would be put into one or multiple category in order to get relevant recommendations with greater performance. This performance would be highly desired especially if user is in the unfamiliar environment with poor-internet connectivity. The impatience for slow-performance could be the experience killer.
The system used a feature vector to compare user models. This two dimension example below explained the idea.
For each user, we create a vector according to the data we collected for each user model.The system used a mathematical rule to transfer each feature in the user model to numeric data. Once we create the vector for each user model, the system compare the direction and magnitude of each vector to decide whether the user model are similar to each other. In the example below, we use the age and average price of restaurants to build vectors.
Wireframes
Design Solutions
The UI design follows most of TripAdvisor branding and design patterns, but adding additional feature to enabling greater end-to-end experience for individual travelers who just wants to get personalized, in-destination quick recommendations based on their preference.
