Erin Shellman - Data Scientists at Work - page 75

The beauty replenishment emails were the first campaign personalized at the
individual level that Nordstrom had launched, and they out-performed
traditional marketing emails. This summer we had a very capable intern
who picked up this work and automated everything into an ongoing weekly
campaign. It's been really rewarding to see that work go from initial idea, to test,
to automated, and now in production. I'm really proud that it's a Nordstrom
product conceived and built in-house, and think it showcases Nordstrom's
technology capabilities.
Gutierrez: How do you or the labs think about whether to undertake new
projects or improve projects you've already done?
Shellman: We usually just have a simple cost-benefit conversation with the
person asking for new features. It typically goes something like this: “I could
build a new recommendation strategy that reads the mind of the customer
and automatically ships products to their house, but I would have to stop
working on the other features you requested.” Not always, but most of the
time the requested improvements are nice to have, but not essential, although
a mind-reading recommender would probably be a good idea to pursue. We
prioritize based mostly on the web teams' launch dates, so as long as we can
hit those we have no problems adding new features.
Gutierrez: How do you think about whether you're solving the right
problem?
Shellman: I typically start from the finish line. Assume that you've built the
thing you're considering, then ask “so what?” Do customers want your prod-
uct and can they use it? Customers can be internal or external. Once we were
considering making a recommendation strategy involving perfume “notes.” In
the end we decided against it because we asked a few simple questions. How
many people use the perfume filters on the website, and who would consume
the recommendations? It turns out most shoppers don't filter search results
by perfume note, and there was no enthusiastic consumer waiting on the
feature so we didn't build it.
Gutierrez: How do get to know the data once you've decided that the project
is worth working on?
Shellman: I spend a lot of time doing basic plotting and visually scanning new
data sets. Our director, Jason Gowans, laughs at me because I'll open and full-
screen a dataset on my big monitor and just scroll through. I just like seeing it
and getting an idea of what the fields are and what it feels like. It's also a good
way to find typos and the weird characters that are common in brand names
like Lancôme and M•A•C.
Plotting data in scatterplot matrices with R is really helpful because you can
quickly start discovering relationships in your data without much work on
your end. Heat maps are great for that, too. I use plyr and dplyr a ton to bust