A Methodology for Discovering how to Adaptively Personalize to User Subgroups using Experimental Comparisons Joseph Jay Williams (
[email protected], HarvardX, Harvard), Neil Heffernan (Worcester Polytechnic Institute). Paper is at www.josephjaywillams.com/mooclet
AdapComp Framework: Experimentally Compare Micro-Designs of Alternative Versions to discover rules for Adaptive Personalization
Abstract We explain and provide examples of a methodology for discovering how to adapt and personalize technology by using randomized experiments to compare alternative micro-designs for technology.
Maximize Target Variable: Response Rate from HarvardX students
The AdapComp Formalism [11] captures an equivalence between using technology to experiment with and personalize modular components of a technology. This provides a simple software design pattern for modular technology components that will allow the use of A/B experiments to discover how to adaptively personalize.
3 Versions of Subject Lines
Any modular technology component expressed in terms of the AdapComp Formalism can be optimized using a general-purpose algorithm that automatically trades off exploration (A/B Experimentation) and exploitation (adaptive personalization to help users).
AdapComp
Goals • •
Are there broadly applicable ways of using A/B experiments to discover rules for adaptively personalizing in real-time? How can live-deployed software be designed to reduce barriers to real-time adaptive personalization?
Contributions • •
Scalable Algorithm for using A/B Experimental Comparisons of different technology components to discover how to adaptively tailor these to user subgroups. Simple Software Design pattern that provides a unified approach to modifying, experimenting with, and adaptively personalizing technology components.
Experimental Variables/Versions 3 Subject Lines x 3 Intro Messages x 3 Response Formats = 27 Versions User Variables: Age Group = [18-22, 23-26, 27-35, >36] Number of Days Active = [0, 1, >2]
Algorithm transitions from A/B Experimentation to Adaptive Personalization: Modify Probability of Versions
ADAPCOMP Formalism User Variable Store
Delivery Rules
TARGET Variable EXPERIMENTAL Variables USER Characteristics
IF [CONDITION] = [A] then show V1 IF [AGE] = [18-22] then show V2 IF [ ] = [ ] then show [ ]
AdapComp Versions
V1
V2
VN
Proportion Students Responding to Emails
Ongoing AdapComp Experimentation & Adaptive Personalization