People problem: Many products suggested irrelevant services to users with Machine Learning (ML), which caused an unsustainable relationship between users and their phones. Our team believed that better-designed intelligent products could serve as a critical shift for phones and develop a personalized relationship with users.

I was the UX Designer on this project. I helped with designing a guideline that specifies when is appropriate for intelligence and how designers should design natural experience with machine learning. I also worked on coming up with use cases and creating prototypes as the best practice of the guideline.

09. 2017 - 03. 2018 @ Huawei Innovation Lab

Does the product needs to be ML-based? How smart it should be?

A framework to evaluate intelligent actions

After knowing the boundary of Intelligent products, I wondered how smart the intelligent product should be. I did competitor research and suggested my peer designers to do an open card sorting activity with me in order to quantify how smart these features are. I divided them into 4 groups and found patterns.

What decides that specific features should stay in a specific state? I first considered two things: the probability of being correct and the cost of being wrong. To better evaluate AI actions, we put those two measurements on the x and y axis. The probability of being correct is heavily correlated to the error rates in machine learning.  And the cost of being wrong takes time, frequency, publicity, monetary, decision ambiguity, and decision fatigue into account. Based on the x value and y value, each action will fall into one quadrant, which is a “state”.

We named the four states:
Manually Maintain: Let users be in full control of the response.

Choice Curation: Present a narrowed-down set of options for the user.

Simplify Selection: Seek user confirmation before execution.

Automatic Assistance: Provide automatic service as well as raising awareness to the user that something was done on their behalf.

I designed some intelligent features that can adapt to four states, and I used those features to validate the framework as use cases.

1. Manually maintain: A user select a song to play.

2. Choice curation: The user gets low-confidence suggestions - more songs from the same artists and the played songs.

3. Simplify selection: With some usages, a medium confidence card queues understand the music genre and habits of the user and queues contextual playlists.

4. Automatic assistance: With more usage, the user gets a high-confidence suggested songs automatically played.

1. Manually maintain: A user search for a destination.

2. Choice curation: The user gets low-confidence suggestions - more destinations from past search.

3. Simplify selection: With some usages, a medium confidence card queues understand user habits of visits, such as time and locations.

4. Automatic assistance: With more usage, the user gets a high-confidence suggested destination and automatically starts the navigation.