4. (Some of the)
Product
Machine
Learning
Problems
Price Accuracy
Ensuring that what you see is what you’ll get
Search
Finding the best itinerary for your needs
Recommendation
Inspiring you to travel to new places
Ad relevance
Connecting partners with the right travellers
Conversations
Go and try our Facebook bot J
Alerting
Keeping you informed, finding the best time to buy
5.
6.
7.
8. Can we do
better?
Historical price focus
Price is only one feature that could make a destination attractive.
Sparse user data
Travel is (relatively) low frequency. Many new, anonymous users –
cold start problem in recommendation.
Destinations are relative
London from Edinburgh is not the same as London from NewYork.
9. …with specific
challenges
No collaborative filtering (yet)
Traditional collaborative filtering algorithms are not suitable for the
data that we have.
No manual intervention
Many approaches that tackle cold-start require manual intervention
from users: profiles, surveys, tags, preferences.
No offline evaluation (yet)
Without data, we have no robust approaches to estimating the
accuracy of recommendations offline (e.g., RMSE).
15. Write the code: The architecture behind Skyscanner’s
recommended destinations (by @AndreBarbosa88)
https://medium.com/towards-data-science/write-the-code-
f6d58c728df0
Initial
Structure
16. Many ways to
define three key
concepts
Popular
Where do people want to (always, recently) go?
“Localised”
What is in higher demand where you are?
Destination-frequency, inverse global frequency.
Trending
Temporal shifts in search behaviours to capture
seasonality, events, demand.
17. Experiments
“Design like you’re right, test like you’re wrong” by @MCFRL
http://codevoyagers.com/2016/03/16/design-like-youre-right-test-like-youre-
wrong/