Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Modeling the Importance of Flight Partners at Skyscanner

techsessions
February 14, 2018

Modeling the Importance of Flight Partners at Skyscanner

Tatia Engelmore, Data Science Manager, Skyscanner

techsessions

February 14, 2018
Tweet

More Decks by techsessions

Other Decks in Technology

Transcript

  1. Modeling the Effect of Removing a Partner Tatia Engelmore Data

    Science Manager @ Skyscanner [email protected] Tech talks, February 8th 2018 1
  2. Skyscanner Partners Skyscanner works with a large number of flight

    suppliers. These include airlines and online travel agents. Sometimes partner connections go down and we can’t pull quotes from them. How big of a problem is this?
  3. Skyscanner Partners A partner connection goes down . . .

    Do we need to wake up an engineer to fix it??
  4. Step back: Why this project? Prioritizing Data Science Projects •

    Impact • Easy to put into production? • Reusability across our verticals (flights, hotels, car hire)
  5. What happens when we remove a partner? Currently we show

    all possible partners for a given route. Is it crucial we show each partner?
  6. What happens when we remove a partner? If Opodo were

    no longer a partner, would people still book on eDreams?
  7. What happens when a partner is removed? Can travelers still

    find all the relevant flights without a given partner? Is the price still competitive without that partner? How much do we rely on each partner for coverage?
  8. Partner removal: iterative approach Strategy • Does the traveler still

    have intent to book? Or do they decide they’re not interested in that booking panel anymore? This is hard. • Simplest model: assume traveler will still click without the partner if they clicked before. Figure out which new partner takes the click. New cheapest partner gets the click (this is our baseline). • Simple model: random forest used for click prediction. • Complex model: two-step model. First predict if any partner will still be clicked, then predict which one.
  9. Partner removal: simple approach 0.9 0.6 0.3 0.2 0.1 0.11

    0.03 0.01 0.0001 0.75 0.4 0.35 0.15 0.13 0.05 0.02 0.0001
  10. Partner removal: simple approach results Results and Next Steps •

    Easy to get good results because vast majority of clicks go to the top position. Most important feature: is partner the cheapest. • Verify model: AB test dropping a partner.