I liked this post a lot.<p>I liked that they are doing this with the simplest and most well-understood techniques, matrix factorization+PCA.<p>Though I'm sure they're also trying end-to-end extra-deep programmatic multi-media networks with new kinds of convolutional layers, fancy residual connections, an experimental batchnorm variant, etc... I'd love to see if they squeeze more juice out of that.<p>As we make more of these semantic meaning-spaces like word2vec, or the meaning space of "photos of airbnb rooms", I think we could use some foundational UI design work to navigate them better.
Like, now you can navigate everything on stitchfix through a small 50-d space where every component is nearly orthogonal and can kind of be interpreted ahead of time+labelled.<p>Is "a binary tree in the form of playing N questions" the best way to go through them, or are there other options? Should we allow navigation in more than one dimensions at a time? Could we start with a few landmark, prototypical elements in each dimension? Have the user progressively clamp down the range in components, and show a PCA of the remaining components each time? Would a 3rd dimension (with VR?) help separate more dimensions at a time, giving you 50% more dimensions to show items on?<p>Computers are increasingly understanding our world and so far they understand it through these meaning spaces, so I think this would be incredibly important for the future of UI design and computer interpretability.
It will be interesting to watch as they improve this process. I use Stitch Fix, and I don't feel like it's all that personalized. The problem is that as a consumer you don't see all the products they <i>could</i> have selected for you. Maybe there's three options for male jeans, maybe there's fifty, it's opaque to the user. The result is a little disappointing because as an American consumer my options are nearly unlimited (only my time is constrained).
Engineering is clearly reaching a point where it is and will be much more integrated and coupled with pure and applied mathematics. I think it's more than time to start taking this implicit project seriously by making it explicit. We aught to think back to the foundations of the way we think about programming in the first place, and we should try to build a new mindset which takes the best parts of the hacker spirit and the pure mathematics mindset and create something new, something more fitting to the enviourment we've found ourselves in, where essentially the only limits to what we can produce is our ability to imagine them.
This is a great article, and it reveals some interesting new ideas to test out. However, I wish they included examples that qualitatively demonstrate how effective each improvement was. Some of us don’t have a team of almost 100 high-end data scientists like they do. Since I only have a couple of people, I need to always be careful to focus on techniques with ROI (low implementation cost, (relatively) high effectiveness). Machine learning can be labor intensive.
Hmm.<p>So I have mixed feelings about this. I think it is good that StichFix is publishing how they work, and I have complete admiration for what they have achieved.<p>I <i>still</i> use their post about the ability to add a "pregnant" vector to an item of clothing and find a similar item in a maternity version to explain how objects can be embedded in a vector space.<p>But this post seems both too complicated for the general reader, and not deep enough for a paper.