Slide 1

Slide 1 text

Constructing Itineraries with Machine Learning @karamshuk

Slide 2

Slide 2 text

Skyscanner Mashups Virgin Delta Potentially cheaper itineraries in more than half of all search results

Slide 3

Slide 3 text

Example

Slide 4

Slide 4 text

Main Difficulty with Mashups • Combinatorial explosion of variants • Cost of constructing mashups is proportional to the volume of variants

Slide 5

Slide 5 text

Start Small Virgin Delta Sum of One Ways • fly-in with one airline, return with another • less combinations, we can effectively sample the ground truth space

Slide 6

Slide 6 text

Supervised Learning Classify whether for a query Q a combination of partners (X and Y) is going to be in Top-10 cheapest

Slide 7

Slide 7 text

Results Coverage or Recall: How many of all possible cheap Mashups our method recalls Cost or Quote Requests: How much quote requests we require for that

Slide 8

Slide 8 text

Temporal Validity Performance of the model stales, hence needs to be updated regularly

Slide 9

Slide 9 text

Production Pipeline Work in progress…

Slide 10

Slide 10 text

Temporal Stability (Origin, Destination, Provider) rules We need a mechanism to control temporal stability of the model

Slide 11

Slide 11 text

Join our Team! [email protected] on Twitter: @karamshuk @SkyscannerEng