Machine Learning (ML) is a powerful tool that helps systems discover and learn patterns of user behavior. This happens when systems are fed data about people’s tastes, their preferences, and their interactions with various platforms. One well-known example is the system used by Netflix, which uses ML to recommend the best movies based on what their subscribers have already watched.
But how does machine learning influence user experience? What can we do to accelerate innovation in this field without losing sight of our main, UX-based objective? Those are tough questions but luckily, Google Design gives us some answers.
Here are 7 things designers need to remember when working with machine learning, according to the folks at Google Design.
1. You Still Have to do the Homework to Discover User Needs
Although ML is a powerful tool, it won’t be of much use if you don’t know anything about the needs of your users. It’s entirely possible to over-orchestrate things, ending up with a powerful ML system for a problem that’s easily solved without it, or even a problem that doesn’t exist.
Before investing in technology, it’s important to know your user and their needs. Otherwise, you won’t know what problem you need to solve. That’s just basic UX, of course, but don’t lose sight of it just because you’re working with ML.
You can go about user research any way you please, from conducting interviews and surveys to analyzing help desk tickets and more. Only by first understanding the user and knowing what their problems are, can you hope to improve their experience through ML. In other words, the first step in the dev process hasn’t changed. It’s to put in the time to find out about the user and their needs.
2. ML Might Not Really Be the Solution Your Users Need
Once you identify the problems or needs of your users, you’ll need to consider everything carefully to see if ML can actually help with a solution.
The challenge at this point is to determine which of the following you’re dealing with:
- Experiences that require ML
- Experiences which will be improved by ML
- Experiences which will be degraded by the use of ML
Not all experiences will require the power of Machine Learning because there are processes that users prefer to take care of themselves.
In order to understand what tasks may require ML, here’s a three-step mini exercise:
- Describe how your user would best perform a task
- If your user performed this task again, imagine how you could simplify things for them
- Describe what could be assumed by the user if a human were on the other end rather than a machine
This should help you to know what assumptions people have when they’re working on a project involving ML.
A good example of a task that doesn’t require ML is email attachment reminders. When the user is going to send an email but forgets to attach a file, a pop-up informs them that they haven’t attached anything yet. The pop-up is triggered when the word “Attached” or “Attachment” is included in the body of the email.
This does not require ML but it helps us to understand the importance of knowing the path the user takes to carry out a task.
For this activity we recommend making a User Journey Map or a detailed User Flow.
3. Prototyping is Possible, and There are 2 Ways to Do It
Another interesting challenge is prototyping a system with ML. If all your evaluations require that you use unique information extracted from a user’s experience, it is practically impossible to quickly generate a functional and authentic prototype. Also, if you wait until your prototype is working in a real way, it’s probably too late to go back and make changes.
To solve this dilemma, you can use either of two methods to obtain the information you need:
- Research based on user experience
- Research obtained through Wizard of OZ studies
Prototyping Based on User Experience
In order to collect useful information from users, you’ll need to make full use of the research stage. It’s important to know as much as possible about your user participants’ daily activities, from the people they communicate with to the movies they recommend. For this, you could perform a simple survey that users could complete as “homework”.
With this information in hand, you could then simulate different scenarios and evaluate the response that the system provides. For example, you could simulate that the system gives the wrong recommendation to the user. This would allow you to study the reactions of the users and how the system should resolve things.
In this way, you’ll be able to more accurately evaluate the cost and benefits of the various possibilities.
Prototyping Through Wizard of Oz Studies
The second approach to gathering research data is called ‘Wizard of Oz studies’. Here, your users are interacting with a ML system that’s actually controlled by one of your developers. You’ll be observing their reactions and their behavior as if they were interacting with a real AI system.
The best thing is that you can do, of course, is to conduct research on your own users.
4. ML Errors Will Affect UX and You Need to Know How
Your ML system can make mistakes and it is important to understand what impression those mistakes will make on the user. You’ll also need to figure out how to prevent those mistakes from degrading the UX.
That’s no easy task. You’ll have to carefully analyze the types of errors that you want to allow into your data set. Keep in mind that, while all errors are the same to a system, not all are the same in the eyes of the user.
For example, if the system classifies a user as a “Troll”, the system won’t know that it is being insulting and that the user will probably be offended and leave, never to return. In other words, false positives and false negatives affect the user and the ML system in different ways. It’s your job to understand how.
To decrease the impact of these types of errors, we must use the right blend of two opposing forces: “remembering” and “precision”. Here’s how that works:
- Remembering. If we aim to “remember” all the errors, the system can increase its effectiveness. The trade-off becomes: we run the risk that those errors will be repeated for a long time until the system recognizes them, learns them, and makes improvements.
- Precision. If we aim at the “precision” of a response, we’re cutting back on recalling many of the errors the system makes. That may help to minimize wrong answers, but it also means you’re missing out on data that might prove important, excluding some correct (yet more obscure) answers. And, by leaving out some of the ‘almost right’ answers, you’re missing out on valuable user feedback. As a result, we might end up providing what the system considers the best response, but perhaps that is not what the user is looking for.
The correct ‘blend’ of precision and recall depends on the problem you’re trying to solve, of course. The cost of mistakes should be prioritized accordingly.
5. A Co-learning and Adaptation Plan is Essential
An ML system will become more valuable as it evolves in conjunction with the user. If the user’s tastes change, the system should learn from that to grant new responses.
While any ML system starts out with existing information, the idea is that, as it develops, it will begin suggesting new options to the user. If the suggestions are not to the user’s liking, then they should be discarded by the user. It’s useful here to implement a survey that’s masked as a query, and plan on getting it out to as many users as possible.
You’ll need to plan for longitudinal research, too. Obtaining the user’s feedback during each session of use is fundamental, in order to capture any changes that may occur in their tastes and preferences. That allows for constantly updating and improving the data set over time for a better ML system.
6. Your Algorithm Will Fail Without Proper Labeling
It is very important to properly label the content you’re working with when you’re creating your algorithms. The danger here is that ML can easily become biased as a result of bad architecture and/or disorganization of the information that’s provided to the user.
The result would be that your ML algorithm begins to collect erroneous information from the user, thus spoiling your algorithm.
The key to success in this regard lies with the users you’ve chosen to help teach the algorithm. Always take into account what and how you communicate with them. For example, use terms that they know and always be sure to assess their understanding. Strategies like Card Sorting can be useful in this context.
7. ML is Collaborative So Don’t Micromanage
ML is an area that involves all parties in the development of the product, so good teamwork is essential. And like with any project involving collaboration, it’s counter-productive to micromanage your people. Nothing stifles creativity more than constantly picking apart every move your team members make.
Currently, the existing tools for ML development are still in their primitive forms. Training your algorithm takes patience, creativity, and a working environment where experimentation is encouraged. Through all this, it’s essential to take into account that if there are limitations on the development side, that’s going to affect the user. Help them out in every way you can.
Human-Centered Machine Learning
Machine Learning is set to grow, as it becomes integrated into an increasing array of products. Our job as developers is to make sure UX is always top-of-mind, no matter what the project calls for down the line. There’s even a name for all this: ‘human-centered’ machine learning (HCML). These seven rules should help you stay focused on the user, even as your jobs evolve to include more machine learning… and they will!
Stay tuned! In our next post we will dig deeper into the fascinating world of Machine Learning, Artificial Intelligence and how they both influence User Experience.