2. 2
George Herber
Project Manager, Analytics at TopBI
Rob Winters
Head of BI, TravelBird
Today’s presentersBuilding a real personalised offer
Together we bring 50
year’s experience across
BI, analytics, and data
science
3. 3
700K emails sent to
NL recipients (3,5M
across EU)
Every person gets the
same 6 offers
Goal: 100%
personalisation
TravelBird: Six daily dealsBuilding a real personalised offer
4. 4
Three objectives for personalisation
Deliver what would someone be interested in
Ensure the right amount of diversity and “freshness”
Send the selection at the most relevant time
Building a real personalised offer
6. 6
TravelBird’s indicators of interest
Pageviews
Email opens
Sales flow interactions
Favorites
Searches
Image clicks
….
Customer Interactions
>500M events over 2,5 years
(but now >15M/day!)
Other Attributes
Similar customers
Time since last activity
User seasonal preferences
“Normal” behaviour
All of this is used to create a score per
customer per offer interaction
Building a real personalised offer
7. 7
Fed into collaborative filtering (like Netflix)
Based on all customers and all products ever, rank online* offers from best to worst for each recipient
Building a real personalised offer
8. 8
Recommendations will be quite similar
Denmark
Germany
Long-haul trips:
(Cuba, Nepal, USA, Iceland,
Morocco)
Building a real personalised offer
10. 10
Building a real personalised offer
And don’t models and data influence each other?
11. Region Similarity: 80%
distance 397 km
Package Similarity: 100%
both incl. flight & hotel,
2/3/4 nights available
Price Similarity: 96%
10 Euro difference
In addition: Text description, image, clicks
Overal Similarity: 96%
So we diversify: Offer similarity
Distance metrics: Canberra, Cosine, Great-
circle, …
Building a real personalised offer
13. 13
In the end: What we builtBuilding a real personalised offer
Events
Monitoring every platform for
user interaction, each day’s
events are fed back into our
databases for inclusion in the
next day’s selections
Models
In Apache Spark we use a variety
of models to come up with
scores for product
recommendations
Diversification
These scores are the enriched
with weather, seasonality, and
other data to build an optimised
planning calendar for each
recipient
Communication
Communication is automatically
scheduled to deliver this
optimised content at the right
time and frequency for each
customer
Delivering +12% opens, +45% clicks, and +22% profit