Machine learning to understand closet info and clothing needs

Machine Learning is going to change fashion ecommerce, and solve real problems people have today. At Chicisimo, we use Machine Learning to understand closet info and clothing needs, and build services on top of that understanding.

We believe that a machine will be able to understand our fashion needs. How can we build such a system?

Chicisimo’s main asset is its Social Fashion Graph, a mechanism to classify any type of input (expression of taste) and to capture correlations among inputs (how needs, outfits and people interrelate). On top of this clean and correlated data, we have built a search engine for outfit ideas, recommendation technology and we’ll be able to personalize content based on personal characteristics.

The backbone of the Social Fashion Graph is our ontology, a game changing classification system that learns from the real world.

The Social Fashion Graph, our apps and algorithms, are the founding patented technology assets towards our goal of automating online fashion services.

The evolution of the music sector can explain the future of fashion

The way we enjoy our clothes and find new ones is going to change. The fundamentals of this change can be explained by looking at the music sector.

In the old days, we used radiocassettes to listen to music. We did not have an online service that could understand our taste and help us find new music. “Discovery” was limited. Years later, Audioscrobbler was born and built a mechanism for us to express our taste, by tracking the songs we played. These data we started to produce was then matched against online databases of songs that were also being built then. This new data allowed product builders to create new discovery experiences, giving birth to online music as we know it today.

Availability of people’s taste, and classification of content. At Chicisimo, we love how the future looks on top of this new understanding.

We are creating a high-quality, vertical dataset, by digitizing offline data

Fashion ecommerce faces a challenge similar to music a decade ago. In order to build game-changing tools, the fashion industry needs to understand people, and not just build new algorithms. And the data required to understand people is, today, mostly offline. What I have in my closet, what I am wearing now, what my context is.

At Chicisimo, our effort consists on digitizing this data, and making it available online. Developing a high-quality dataset is the success factor towards building a human-level tool to offer outfit advice. Data is a critical element in Machine Learning for fashion.

Chicisimo is a learning mechanism. And we love how the future looks on top of this new understanding.

The Social Fashion Graph

The objective of the Social Fashion Graph is to provide structure to the data generated by people. And the objective of this structured data is to enable automated and personalized services.

The Social Fashion Graph has two fundamental tasks: (i) classify input: it converts any expression of taste into structured data, regardless of how the input was expressed; (ii) capture correlations among data: it captures how needs, outfits and people interrelate.

As a result of the above, the Social Fashion Graph produces clean and correlated data, a well curated and growing fashion dataset that learns from the real world. This structured data is an enabler of services. As of today, we have built a search engine for outfit ideas, recommendation technology and we’ll be able to personalize content based on personal characteristics.

The backbone of the Social Fashion Graph is our ontology of the world’s what-to-wear needs. It summarizes and gives structure to fashion needs, as expressed by people.

We think the ontology is pretty unique, and here is why: The traditional fashion taxonomies are a list of tags describing garment characteristics. This approach is understandable, because it is looking at clothes through the lense of other products it has worked on before (songs, books, electronics…). These other lines of products had pre-defined databases, obvious matching systems, and a type of metadata much closer to the way people describe those products.

We have learnt that fashion needs go well beyond garment-related metadata, and of course these needs are expressed in very different ways. In the age of deep learning and substancial algorithmic advances, next generation fashion taxonomies are transparent and capture the world through the eyes of people, not through the eyes of the industry.

This social, bottom-up approach contributes to smarter algorithms. It also expands the discovery experience: it empowers fashion shoppers to look at each other to decide what to wear and what to buy, instead of having to look at the traditional gatekeepers.

“I need ideas to wear this violet sweater to class tomorrow”

Did you know that most people combine violet sweaters with black pants, black boots or black blazers? Also, violet sweaters are more common during fall/winter than during spring/summer. And they tend to be worn during casual settings rather than formal.

Data is cool. But being able to recommend the right outfit for a violet sweater requires a deeper understanding of what someone likes and needs. It also requires creating the right environment for people to express their needs, where inspiration and discovery can take place.

The above is specifically what we are building.

Thanks for creating this fashion app, it helps me enjoy my clothes, great outfit planner
I love fashion, and feeling pretty makes me feel stronger. This wardrobe app helps me getting there
I totally love helping other women with my selection of fashion ideas

Product first

Product first. Building the right product experience is first for us.

It also allows us to learn what is the data that makes an impact and what is the technology that needs to be built, in order achieve our objetive of automating outfit advise.

A few examples: Thinking in terms of product and retention helps us focus on creating user habits. Thinking about the onboarding process helps us overcome the cold-start problem. Understanding product cognitive overhead, leads us to search a specific type of design.

And we love the feedback we receive, and how we are regularly featured as App of the Day by Apple in more than 60 countries in 2017, and being Android’s Best App in Spain and France 2015 and 2016.

Please visit this essay to learn about the process we followed to built our product and data platform.

Fashion. 10 years from now

Brands and clothes have immense meaning, because people express through them. This meaning can now be taken into account. Machine learning is the enabler, and it will change fashion ecommerce. A decade ago, we did not have Spotify. A decade from now, we will have several “Spotify for fashion”. There are some specific data-based use cases we like, that fit well together:

  • Automating outfit adviceAmazon, Alibaba, Chicisimo
  • Image understanding (Most ML efforts are here)Wide-Eyes, Vue.ai, ViSenze
  • Systems to match outfits and shoppable productsInstagram, Pinterest, Amazon, Chicisimo
  • Discovery of clothesAmazon, Zalando, Asos, Chicisimo

If the above makes sense to you, please

We use our own and third party cookies to offer you a better user experience. When you use our services, we understand that you accept the use we make of cookies.