a practical framework of conversion rate prediction for
play

A Practical Framework of Conversion Rate Prediction for Online - PowerPoint PPT Presentation

A Practical Framework of Conversion Rate Prediction for Online Display Advertising Quan Lu, Shengjun Pan, Junwei Liang Wang, Hongxia Pan, Fengdan Wan Yang Alibaba Group. CPA vs CPC Trend Number of publications at each Percentage of


  1. A Practical Framework of Conversion Rate Prediction for Online Display Advertising Quan Lu, Shengjun Pan, Junwei Liang Wang, Hongxia Pan, Fengdan Wan Yang Alibaba Group.

  2. CPA vs CPC Trend Number of publications at each Percentage of campaigns with each year goal type 2

  3. Rarity Extremely low conversion rate. Typically in 10E-5 to 10E-6 range. Attribution Heterogeneity There is no 100% causal Learning across different relationship between the advertisers is not allowed. events and following conversions Delayed Feedback Challenges for CPA Days or weeks delay before getting the response. Predictions Different types Post-click-action and post-view-action are naturally different in modeling.

  4. Challenges of Conversion Predictions Due to Conversion Rarity Worse performance u Low delivery u Costly exploration u 4

  5. CVR Safe Prediction Framework 3 2 1 Growing stage New exploration regions are added based on the cross region conversions to provide continuously explorations. Evolution stage Each region’s predicted CVR is converted to its own empirical mean. So, some regions’ outputs automatically fade out. Initial stage Exploration regions are derived from the trained GBDT models. Each region may have different CVR priors.

  6. Performance Comparison

  7. Conversion Adjustment for Delayed Feedbacks Trade-off of including new data with u compensating empirical estimation. P4 Formulate and solve as the following u P3 linear programming problem. P2 P1 Current Time T Conversions Adjusted Conversions at T-1 = N1 / P1 7

  8. Bid Price Adjustment for Last-Win-All Attribution For post-view-conversion campaign, u impression value decreases when user has been shown the same impression before. The discounted value positively correlates u to the elapsed time between the last shown impression at the same user The key is to conduct noise reduction and u do fast approximation of the integration value online 8

  9. Over-prediction Gap Brought by RTB Gaps between observations u and predictions could be partially introduced by real-time Winning Impressions bidding. Bid Predicted CVR Price For CPA remarketing GAP u impressions, this phenomenon Observed CVR is more obvious due to sharing common information 8 Impressions with the same predicted CVR 9

  10. Problems are not stop signs, they are guidelines - Robert H. Schuller 10

Recommend


More recommend