Cost-sensitive Learning for Utility Optimization in Online - - PowerPoint PPT Presentation

▶
cost sensitive learning for utility optimization in
SMART_READER_LITE
LIVE PREVIEW

Cost-sensitive Learning for Utility Optimization in Online - - PowerPoint PPT Presentation

Sep 02, 2022 483 likes •717 views

Cost-sensitive Learning for Utility Optimization in Online Advertising Auctions Flavian Vasile (Criteo) Damien Lefortier (Facebook) Olivier Chapelle (Google) Agenda Context Online & Offline Metrics Utility Optimization

slide-1
SLIDE 1

Flavian Vasile (Criteo) Damien Lefortier (Facebook) Olivier Chapelle (Google)

Cost-sensitive Learning for Utility Optimization in Online Advertising Auctions

slide-2
SLIDE 2

Agenda

  • Context
  • Online & Offline Metrics
  • Utility Optimization
  • Online & Offline Results
slide-3
SLIDE 3

Context (1)

  • Online Advertising Auctions for Display Advertising; 4 types of players:
  • The auction house: RTB platform,
  • Demand: the advertiser,
  • Supply: the publisher,
  • Bidder: Criteo.
slide-4
SLIDE 4

Context (2)

  • Most people optimize for deep-funnel events and use a conversion

rate (CR) prediction model. We focus on this aspect here.

slide-5
SLIDE 5

Agenda

  • Context
  • Online & Offline Metrics
  • Utility Optimization
  • Online & Offline Results
slide-6
SLIDE 6

Online Metrics

  • Conversions are different (e.g., sock vs. car) so we need to weight

them by (some flavor of) CPA = Cost / #Conversions.

slide-7
SLIDE 7

Offline Metrics

slide-8
SLIDE 8

Agenda

  • Context
  • Online & Offline Metrics
  • Utility Optimization
  • Online & Offline Results
slide-9
SLIDE 9

Utility Loss

  • Defined as the opposite of the Utility:
  • Non-convex; very hard.
slide-10
SLIDE 10

Utility Loss and Log Loss (NLL)

  • We analyze the Utility loss when c is close to our bid pv.
  • We assume conversion probabilities are small (p << 1).

=> The derivatives are approximately equal, up to a factor v

slide-11
SLIDE 11

Toy Example

  • Two advertisers with different CPAs (5 and 50) and CR (1% and 0.1%).
slide-12
SLIDE 12

Method

  • We propose to optimize for WNLL to improve our bidder’s performance.
  • We use L-BFGS for learning.
slide-13
SLIDE 13

Impact on Regularization

  • We propose the following heuristic to take weights into account:
slide-14
SLIDE 14

Agenda

  • Context
  • Online & Offline Metrics
  • Utility Optimization
  • Online & Offline Results
slide-15
SLIDE 15

Offline Setup

  • We use a public Criteo dataset for our experiments.
slide-16
SLIDE 16

Offline Results – Weights

slide-17
SLIDE 17

Offline Results – Lambda

slide-18
SLIDE 18

Offline Results – High/low CPA

slide-19
SLIDE 19

Online Results

  • The A/B test was done on more than 1 Billion ad displays, on world-

wide traffic. Our change resulted in a +2% lift in ROI.

  • We observed significant savings in display cost + an increase in sales

performance for the advertisers, especially on the campaigns with high CPA and low number of sales.

slide-20
SLIDE 20

Conclusion

  • Weighted Log Loss allows to get closer to both offline and online

metrics in the context of online advertising auctions.

slide-21
SLIDE 21

Thanks!