How we grew from 0 to 4M women on our app, with vertical machine learning(medium.com)
medium.com
How we grew from 0 to 4M women on our app, with vertical machine learning
https://medium.com/@aldamiz/how-we-grew-from-0-to-4-million-women-on-our-fashion-app-with-a-vertical-machine-learning-approach-f8b7fc0a89d7
20 comments
Thanks for the heads up, dang.
Truth is: for the first time ever, we've decided to talk about our machine learning approach and how we've managed to build our project, and we've thought many of the lessons learnt have value for the community.
We've really worked hard on the content of the post, trying to offer valuable insight on a number of things: how we work retention, onboarding, what's our learning process, and most importantly how we understand fashion taste. Lots of tips that imo are valuable for the community. And all this work, we've wanted to share it here on HN. Our friends are as excited as we are, and some have asked for questions / congrat'ed us here. No questions were planned.
About the title, it's what we've managed to do! While others spend millions on acquisition, we haven't, it's been product based, no tricks, its been done by building an efficient product, step by step, countless nights. Few people in this industry, outside the fashion space, know Chicisimo, and we wanted this to come to an end. I'd appreciate the flag to be lifted.
Truth is: for the first time ever, we've decided to talk about our machine learning approach and how we've managed to build our project, and we've thought many of the lessons learnt have value for the community.
We've really worked hard on the content of the post, trying to offer valuable insight on a number of things: how we work retention, onboarding, what's our learning process, and most importantly how we understand fashion taste. Lots of tips that imo are valuable for the community. And all this work, we've wanted to share it here on HN. Our friends are as excited as we are, and some have asked for questions / congrat'ed us here. No questions were planned.
About the title, it's what we've managed to do! While others spend millions on acquisition, we haven't, it's been product based, no tricks, its been done by building an efficient product, step by step, countless nights. Few people in this industry, outside the fashion space, know Chicisimo, and we wanted this to come to an end. I'd appreciate the flag to be lifted.
Is all the tagging manual?
I assume that in order to find images of relevant outfits for the expressed "needs", the images have to be tagged with colors, brands, garment types... If that tagging is manual, I wonder if it could be automated using the object detection feature of an image analysis service like Amazon Rekognition or Google Cloud Vision.
Maybe automated tagging would allow richer tagging, and that could be key to find the best results for each user's taste.
I assume that in order to find images of relevant outfits for the expressed "needs", the images have to be tagged with colors, brands, garment types... If that tagging is manual, I wonder if it could be automated using the object detection feature of an image analysis service like Amazon Rekognition or Google Cloud Vision.
Maybe automated tagging would allow richer tagging, and that could be key to find the best results for each user's taste.
How much have you spend on Mixpanel??? ;)
How did you get to insights/assumptions like "If people don’t do [action] during their first 7 minutes in their first session, they will not come back"?
How did you get to insights/assumptions like "If people don’t do [action] during their first 7 minutes in their first session, they will not come back"?
Thanks for the question furilo.
We run behavioral cohorts. We do cohorts over everything, every action people do, every screen they visit, every piece of value they receive. We also know when those actions take place (first 5 minutes, first hour, day, etc).
With the above, we've learnt what actions have the strongest retention. Now we need to look at the duration of the first session, which we have. We know that if we dont convert people during their first session, they is no coming back - no push or email or anything, will bring them back.
Once we know what the lever of retention is/are, and much time we have to convert, we run user tests with different types of users (converted, non-converted, people who've abandoned the app, people who dont know the app), and observe how they discover the lever, how they describe it, how they use it. This is gold, and helps us iterate.
We run behavioral cohorts. We do cohorts over everything, every action people do, every screen they visit, every piece of value they receive. We also know when those actions take place (first 5 minutes, first hour, day, etc).
With the above, we've learnt what actions have the strongest retention. Now we need to look at the duration of the first session, which we have. We know that if we dont convert people during their first session, they is no coming back - no push or email or anything, will bring them back.
Once we know what the lever of retention is/are, and much time we have to convert, we run user tests with different types of users (converted, non-converted, people who've abandoned the app, people who dont know the app), and observe how they discover the lever, how they describe it, how they use it. This is gold, and helps us iterate.
Which action has the strongest retention in a fashion app like yours?
Article suggests they are using perceived value of an outfit to train their Machine Learning model. I wonder what are those actions, Likes? Add to cart?
Thanks for the comment soci.
There is a lot of actions that express taste - sharing your own tagged outfits, searching for something, creating a specific album, and many more. And actually, each expression of taste has its own value. As we move forward, we'll need to get smarter as how we can help people, based on what data.
There is a lot of actions that express taste - sharing your own tagged outfits, searching for something, creating a specific album, and many more. And actually, each expression of taste has its own value. As we move forward, we'll need to get smarter as how we can help people, based on what data.
Fantastic article and way to go on the accomplishments. Looks like you are in an amazing space to be. I am a big big fun of recomender systems.
thanks quarizmi!
Lets see how recommender systems evolve in this space. The tech and implementation will need to be very different to traditional recommenders, and access to data will be a huge barrier imho. We'll see:)
Lets see how recommender systems evolve in this space. The tech and implementation will need to be very different to traditional recommenders, and access to data will be a huge barrier imho. We'll see:)
Thanks for sharing! You’ve talked about your ontology, can you explain a bit more about how you build it?
Thanks for the question.
The question we figured out we need to answer is: what type of outfit ideas are people looking for, and how do they describe their needs? And then, how can we match those needs with relevant content.
So what we did (what we do) is look at people's queries, and also at how people describe their clothes, their outfits or albums (combinations of outfits, like pinterest boards). There is a lot of data here, millions of keywords, but many of them are similar (pants and trousers, or pantalones in Spanish). So we’ve extracted the main concepts (pants) and built a system of equivalences. Basically, now, we know “all” the needs people have when it comes to deciding what to wear, and have a pretty complete view of different ways of describing those needs, and how good the system is at responding to those needs (database of outfits).
The question we figured out we need to answer is: what type of outfit ideas are people looking for, and how do they describe their needs? And then, how can we match those needs with relevant content.
So what we did (what we do) is look at people's queries, and also at how people describe their clothes, their outfits or albums (combinations of outfits, like pinterest boards). There is a lot of data here, millions of keywords, but many of them are similar (pants and trousers, or pantalones in Spanish). So we’ve extracted the main concepts (pants) and built a system of equivalences. Basically, now, we know “all” the needs people have when it comes to deciding what to wear, and have a pretty complete view of different ways of describing those needs, and how good the system is at responding to those needs (database of outfits).
Awesome article with tons of learnings and details <3. I'm a big fan of them.
Amazing work!
Since it looks like the underlying work here is good, I suggest waiting a week or two and then reposting it, and make the title less buzzwordy and more neutral. If you email us at [email protected] when the post is up, we can make sure it doesn't get flagged.