7 July 2018

Mailer Campaign Uplift Modeling

Profit(C) = ExpectedProfit(C) x [P(B | V) - P(B | C)] - AdCost(C)

  • P(B | C) - probability of buying given control without ad campaign (Naive Bayes)
  • ExpectedProfit(C) - profit to make from customer if they decide to buy (Regression)
  • P(B | V) - probability of buying given variant of ad campaign (Naive Bayes)
  • AdCost(C) - cost to mail campaign to customer as a constant
  • likely to take into account market or customer segmentation
  • regression could be either logistic or linear
  • total profit would be determined by how much the customer decided to buy either with control and/or ad campaign
  • optimization of ad campaign given the customer conversion ratio
  • use customer data as part of expected profit measures for average spend
  • additionally, more ways to approach the same contextual measures of profit

3 July 2018

Test-Driven Machine Learning

TDD -> Kent Beck
BDD -> Dan North
Refactoring -> Martin Fowler
Agile -> James Shore

Random processes in machine learning need to be measured and controlled, various simple testing strategies can make this possible.