Girish Sathyanarayana, Senior Data Scientist at AppLift, " Business Value Through Data Science: A Mobile Ad Tech Case Study"
2 likes•531 views
The document outlines the operations of a global expert service with over 200 employees across 10 locations, focusing on mobile advertising campaigns through an integrated platform. It discusses the challenges faced by demand-side platforms (DSPs) in leveraging data science for campaign success, emphasizing the importance of model predictions for bid value and the complexities in achieving consistent campaign performance. The text highlights the use of feedback control systems for bid adjustments to meet advertiser objectives while acknowledging the realistic limitations in generalization of models.
Girish Sathyanarayana, Senior Data Scientist at AppLift, " Business Value Through Data Science: A Mobile Ad Tech Case Study"
- 1. 1
- 2. Expert Service 200+ employees from 40+ Nationalities We are in 10 Global Locations: Berlin, Tokyo, Moscow, San Francisco, Seoul, Bangalore, New York, SĂŁo Paulo, Shanghai and Jakarta 2
- 3. DataLift 360 3 Our integrated platform is your single access point for successful mobile advertising campaigns
- 4. 4 Business Value Through Data Science The Problem: If you are a DSP trying to run advertisers’ campaigns, how can you leverage Data Science?
- 5. 5 Opposite Ends in Ad Tech Spectrum Mobile AdvertiserData Scientist
- 6. 6 Different Views • Advertiser (Business View): Running campaigns with a target CPI and scales target. • Data Science: Build an accurate model that makes quality predictions.
- 7. 7 Forging Components • Model predictions used to quantify the worth of an impression (bid value). • How does this Data Science approach meet Business Needs. • Depends on how well model generalizes.
- 8. 8 Ideal Scenario • The perfect model would have good generalization (i.e. performance on unseen data). • Good generalization at the campaign level is important: • Achieve a CPI < Target CPI. • Does not necessarily mean you’ll achieve your scales objective.
- 9. 9 Realistic Scenario • Realistically, it’s difficult to achieve perfect generalization. • Most times, there is no good generalization at the campaign level. • Some campaigns do well, some don’t. • Does not solve the business problem.
- 10. 10 Bid Adjustment • Use model output as central value around which we vary the actual bids. • Vary bids by computing an error for budget delivery. • Basically, a Feedback Control System.
- 11. 11 BID VALUE FROM MODEL BIDDER ERROR PROCESSING BLOCK BID CORRECTION ERROR REFERENCE TARGET CPI & SCALE TARGET REALIZED CPI & SCALES OUTPUT Feedback Control System Possible Outcomes • Advertiser’s demands are met (i.e. target CPI and Scales). • Or algorithm flags that the campaign objectives are not achievable with recommendations for achievable targets.
- 12. 12 Q&A Thank you for listening! Any questions?