Vizit · 2015 to 2018
Image Performance Modeling
Predicting how a specific image performs with a specific audience, so the choice of creative stops being a matter of taste.
Role · Principal Data Scientist.
The problem
A brand has one product and a dozen images of it. Which one goes in front of which shopper is decided by whoever has the strongest opinion in the room, and the feedback loop is slow enough that nobody ever finds out they were wrong. The underlying question is not whether an image is good. It is whether this image works on this audience, and those two things come apart constantly.
Approach
- 01
Modeled imagery on visual attributes rather than on metadata, so an image the model had never seen still had a position in the space.
- 02
Scored performance against a defined audience rather than in the abstract, because a single quality score throws away the only part of the answer that is actionable.
- 03
Built the evaluation around held-out performance data, since a creative model that cannot be falsified turns into an opinion with a number attached to it.
Outcome
Held alongside a full-time buy-side research role and the Booth MBA. The problem shape, choosing among assets for an audience, is the one I have kept returning to.