Efficient Data Labeling

Early users complained of the confusing and friction heavy data training experiences after the release of the beta version of We redesigned the process to simplify the data tagging experiences to increase the incentive to tag and enhance the accuracy of AI.


Full-stack designer— discovery, user research, design, testing


A project manager
A developer
2 designers
2 data scientists


Summer 2019, 4 Weeks

Challenges beta version was released as a A.I. training software for learning social media’s posts. Users need to categorize and tag social media posts for AI to learn. The more the tagged data, the more sophisticated the A.I. prediction. However, in the beta version, the tagging process was tedious, so not many users used it.


An Easy A.I. Training Process

Label posts and add new topics become easy

See and confirm A.I. prediction frictionlessly


How we got there...
Our Goal

Success meant a growing number of data labeling, which would increase the accuracy of the AI prediction. Our team’s primary goal was to simplify the data tagging feature so users could easily and quickly train AI to learn social media’s content. In addition to that were higher-level goals for the business: increase the trustworthiness of the software, transform tagging labeling process into a streamlined experience, enable users to correct the AI easily, and help users to understand the AI model they have built.


Tangible to digital

I first began with high-level sketches to get a rough framework of the interface on paper before diving into the details.

After experimenting on paper, I began designing variations in Figma. I would show my team the designs I was most confident about, then get feedback and go back to the design for another iteration. Every single detail was nuanced, and I felt as if I was trying to mix and match all these important pieces together into a jigsaw puzzle.

Early Prototyping

Familiarity: categorization

We drew inspirations from systems that have similar categorization functions such as bookmarks on social media posts, tagging on blog posts, and file labeling in the computer systems. Hence, this could simplify the learning process for users.

Moreover, early users interviews show that users want to know the source and creator of the posts, and the number of reactions to the post in order to grasp the virality of the topic.

Early Prototyping

Exploring the folders’ structure

AI learns from the topics’ tag and its corresponding sub-topic tags. That means users need to find subtopics inside a topic to tag data. I used an arrow to indicate this hierarchy, similar to the folders’ structure in computer systems, arguing that it allowed users to have a clear idea of the relationship between tags.

The feedback reminded me of our goals— simplify the data tagging experiences. The folder structure confused users. Hence, I moved away from the computer file’s structure, as it contradicted with our belief that more data tagging is better.


Handling the complex machine learning system

In supervised machine learning for social media, AI needs to have 3 types of tags and 3 types of sentiments. For AI-predicted tags, users need to confirm if the tags are correctly labeled for further machine learning. There are also 3 levels of sentiments for AI to learn the subtle emotion which every social media post conveys.


Differentiating tags

I explored options to distinguish the different types of tags in various visual styles, which is to minimize the amount of jargon in data science.

Three types of tags

I designed the three types of tags in these three ways, but the feedback and potential users feel that the visual details are too subtle.

Three colors for three sentiments

I put the three sentiments in three colored dots, which is to minimize the amount of wordings. However, we found that the users ignored the circles indicating sentiments when tagging.


Keep streamlining

From our user testing and data, it is very rare that there are the same subtopics’ names in different main topics for one data. Users would include hints of the topics when naming the subtopics. Therefore, I removed the folders' structure, instead, users could hover on the subtopic tag to see which main topic the tag belongs to.

On the other hand, the check mark also confused users about the status of the tag. They could not understand what they could do if they check the box. Hence, I removed the checkbox in the final design.

Narrowing down the types of tag

After discussing with the development team, we understand the distinction between user added tag and user confirmed AI tags does little to help the sophistication of the algorithm. Therefore, we narrow down the types of tags to two.


How to adapt to changing requirements

New timelines, resourcing issues, and reprioritization meant the scope of the project was constantly changing. I had to adapt to those changes and still deliver the best design in time with tight deadlines.

Always fight for good UX

I had to work under very strict technical constraints of machine learning, but still fight for what I believe is essential to having a good user experience.

Previous project

Social Monitor

Next project

Gift Ceremony