How was the Chicisimo dress sharing application that started from 0 and won 4 million users and was developed?

Google is "Style IdeasIn recent years, Amazon, Google and others are entering the field of fashion and dressing, etc., "dressing" takes root on people's everyday, looking at fixed visitors and users It is said that it is because you can get a lot of data from. It is also predicted that the revenue of the apparel market using e-commerce will reach $ 123 billion in 2022. Meanwhile, an application that increased the number of users to 4 million people in a few years starting from 0 "ChicisimoFounder ofGabriel AldamizThinks about how the application development was promoted.

How we grew from 0 to 4 million women on our fashion app, with a vertical machine learning approach
https://medium.com/@aldamiz/how-we-grew-from-0-to-4-million-women-on-our-fashion-app-with-a-vertical-machine-learning-approach-f8b7fc0a89d7

Chicisimo is an application that allows users to post their own fashion coordinated photos and evaluate them from other people, or to get the idea of ​​dress from people. Although it originally started as a blog, the application was released, and as of 2018 there are 4 million users.

Chicisimo - The outfit ideas app to decide what to wear
https://chicisimo.com/


Subtle factors such as "style" and "taste" are involved when a person makes a fashion proposal to someone else. It is possible for humans to understand them and make new proposals from already existing dresses, but in the case of applications, it starts with making computers understand "style" and "taste" first.

Chicisimo's development team has previously incorporated machine learning in the field of music and other projects based on "taste". Based on past experiences, the team thought that if the computer understands "taste", it will be able to provide more relevant and meaningful content, online fashion will change.

Especially the development team focused on making the right dataset and developing two things, "mobile application" and "data platform". Chicisimo 's CEO Gabriel Aldamiz looks back on how development was progressing as follows.

◆ 1: Developing applications that people express their needs
Aldamiz, who learned that it is very difficult to continue to use "even if it is easy to use the application to people" from the experience of past application development Aldamiz, in developing ChicisimoIterationI tried to do it as fast as possible.

Chicisimo got a very early alpha version with only key features first published outside the United States. This application is not named Chicisimo but it was deleted from App Stpre once the official version was released. There are no photos left by the user uploading to the application at the moment. However, thanks to the alpha version, real data for iteration and good input were obtained.

Especially Aldamiz's emphasis was on "making algorithms to match people and content" and "RetentionUnderstanding the elements that enhance ".

According to Aldamiz, there are three things that help to increase retention:

(1) Recognize something that enhances retention by cohort analysis
The team not only said that "what kind of behavior did the user take", but "what was it useful for?"MixpanelWe used a cohort analysis. Although it was a very difficult thing, it seems that we repeatedly analyzed, tested and improved by finding measurable values. And, among them, it also identifies things that worsen retention, and said that they removed them.


(2)On Boarding· Review the process
Aldamiz's onboarding process is "to find the app's value as soon as possible before you lose the user." Aldamiz thinks "If the user does not take action within seven minutes of the first session, they will never come back", and by having people of different types repeat the test many times, the user It seems that he made an application experience to "take action".

(3) Decide how we learn
Data approach is important, but there are things more important than data to make products loved by people. In the case of Chicisimo it was an understanding that "the problem is very important". It is said that it became a way to show respect to people.

As a result of the above efforts, the development team could acquire new knowledge like a mountain, and said that it was able to greatly contribute to product development. When I encounter new knowledge and learning contents that change the flow so far, I will focus on two points, "how people are involved in problems" and "how people relate to products" Need, and Aldamiz. Understanding these two will lead to the development of apps that people love.


When talking to colleagues Aldamiz got the opinion "This is not about data but about human beings", Aldamiz talked directly to women about problems related to dress and dressing and solutions, or e-mail opinions We proceeded the investigation by receiving it. Also, looking to the outside, talking with people who are developing interesting apps, and rereading articles that seem to be helpful were also repeated.

Applications that were made in this manner were successfully attracted by the App Store, etc. By 1 January 2018, the application reached 970,437 views and the conversion rate from impression to installing the application reached 0.5% He said that he did.

◆ 2: Make data platform to learn the needs of people's fashion
Chicisimo's goal is to understand the user's taste and show clothing ideas. If you show the right content at the right time, you can make people say you are saying this is a thing called easiness to go.

Chicisimo is 100%User generated content, The system should automatically classify the content type, create appropriate incentives and understand if the content meets the user's needs. No matter how much data gather, unless the system can properly process the data, it will not be possible to use the information and only chaos will be born.

So the team first developed a tool called Social Fashion Graph. This is to make some parts of one piece of data have a structure. The graph created by Social Fashion Graph visualizes needs, dressing, people's relationships, and says that this concept was useful for building a platform. Social Fashion Graph made it possible to create a high quality data set, and it helped to learn the application.


According to the development team, "Coordination of clothes is music play list"Collaborative filteringIt seems that we were able to offer "Recommended" in various places in the application by catching relevance using.

However, even if Social Fashion Graph is applied, noise still exists in the data. People express one "needs" in different ways, and even if combinations of similar clothes are different needs or vice versa, the same needs may be the same.


This need is "concept" such as "clothes to wear to school" and "weekend fashion". In order to capture the diversity of expression of people, the development team incorporates elements of concept into the system. Then, the same need expressed in different ways was regarded as equivalent, and a list of the needs "what should I wear" eventually was created. I was able to organize the data set by creating this list.

If you can structure three of dressing, needs, and people, you will be able to understand large amounts of data. Even if the user freely expresses it, if the correct system is behind the scenes, the data will be structured and control will be available. Meanwhile, Aldamiz says that unstructured data will give the development team new knowledge and flexibility.

The development team of Chicisimo still has challenges but I feel that they are in their "managed" state. As of February 2018, while trying to add a new element of "Possible clothing" to Social Fashion Graph, this element should help the user to consider "what to buy next" is.

◆ 3: Algorithm
Mr. Aldamiz looks back to it that it was not difficult to make "Recommended" system for music application development. It is easy to grasp the songs that the user likes and it makes it easier for people who likes that music to grasp the next songs that they are likely to hear.


However, as you can imagine if you think of the closet, it is rare for a person to "buy the same clothes as clothes you already have", and you will make a choice like "buy clothes that match the clothes you already have" . Such a sequence is difficult to see the relevance, and it was difficult to create an "recommendation" system. Also, the element "style" is complicated, it seems that it was difficult for computers to catch and classify.

However, with the appearance of deep learning, the circumstances have changed, Aldamiz. Prepare the correct data set, you do not need to gather or scrutinize the data, adjust the finer details related to "Recommend" or to focus on delivering value through the algorithms. The algorithm created by deep learning is still being used secretly by the application, but it is said that we will further improve it based on future feedback.