Recommender system basics: Part 2

Recommender systems present people items that are best guesses for what they may need, like, want to consume or interact with. They define an important part of the user experience of many products. Because of that, designers should be involved when recommenders are created. User research, user-centered goals and evaluation of the impact can be deciding factors for a recommender.

In this post, we will go into more detail about what designers can do exactly in the creation of recommenders. But first a famous anecdote about why designers should be involved in what’s often framed as an engineering challenge.

Reasonable recommendations are relatively easy to create, but great recommendations are hard. So, in 2006, Netflix announced a contest with a prize of $1 million for whoever could make a recommender that performed ten percent better than the existing system. The prize was granted three years later, but the winning algorithm was never implemented because the costs for implementation were so high. Moreover, Netflix found better ways to present recommendation.

It turned out, changes in the UI made a bigger difference in the usefulness of the recommendations than the improved algorithm could.

Define purpose

Perhaps the most important influence we designers can have is in making sure the recommender is created to help users reach their goals. As shown in part 1, that’s not always a given. Designers are often close to users through the research we do. We may be able to better specify such goals to ensure they’re not just good for short term business.

Things to consider:

  • Any list in a UI is an opportunity to show automated recommendations. Sorting items alphabetically or by date is rarely the optimal solution. Keep in mind though, strictly speaking, recommenders are used to present users items they don’t know yet. But in early ideas, that doesn’t matter yet.
  • How can recommenders help users achieve their goals? Does someone shopping for jeans want recommendations for more jeans, other sorts of pants, or for T-shirts? Does the context in which the user goes to a video service matter to what type of video they would like to watch?
  • What type of user needs recommendations the most? Somebody just starting with the product, or a long time user? How are their needs different?

Ultimately, such questions may be answered once recommender systems are in place. When data is collected, algorithms can identify if people on transit want long or short videos, or what recommendations are most effective to novice users. Until then, a working hypothesis can help the recommender system to start, get used and collect data for improvement.

Balance business goals with user needs

Even today, many engineering projects start without design involvement. Recommenders are created to meet business goals with little concern for users. A well-known example: “Recommend content users may find interesting, so they spend more time on our ad-supported platform.”

A former Google engineer now exposes how YouTube recommendations, that are made to maximize watch time, aren’t necessarily beneficial for the user. His AlgoTransparency project shows how YouTube promotes clickbait and extremist views. Taking business goals to the extreme can hurt the user experience and ultimately, the business too.

Identify possible unintended consequences

I presume YouTube didn’t create their recommendation engine with the purpose of serving seductive, low quality videos. It just turned out that that’s the type of content that surfaces in the survival of the fittest competition of video publishers.

When ideating on ideas for recommender systems, we should consider how they can be manipulated or have other negative consequences. We like to think positively about the things we create, but what could be negative effects of a recommender? Movie recommendations may seem innocuous, but what if certain recommendations can be linked to teenage suicide rates going up?

The Ethics Center at Santa Clara University created seven tools you can use on any type of design project. They can help to steer your product away from unintended negative effects and improve the user experience.

Present recommendations well

Before a recommender system is built, the product team should have a clear view on how the recommendations will be presented to the users. That’s not your typical UI design job though. Typically, design prototypes can only show manually created recommendations. It’s hard to take a participant’s preferences into account there. As the usefulness of the recommender system depends on the quality of the recommendations, such a prototype has limited use.

The team can still learn something from tests with such prototypes, like what part of the user flow is good for doing recommendations. Recommendations can be perceived as salesy and obtrusive, but they can also help undecided users find what they need.

Perhaps the most important choice for presenting recommendations is whether they’re shown as many or one. A list is the obvious first choice: it’s just more likely that a user finds something in a list of ten than when there’s only one option. It’s even more likely when that list is longer.

Present many items

Recommendation lists are most effective when they have more than a few items, so think of how much space a list needs and if it can be scrolled or expanded. Keep in mind that creating a long list of recommendation can take a while. If they’re created on the fly, generating many may take a few seconds. By the time they have been loaded in the UI, the user may already have scrolled past them.

Even better than one massive list are multiple types of recommendations presented simultaneously. That way users can look for something based on what they at that moment find relevant. Amazon’s home page is an example of that. After you log in, it’s filled with all sorts of recommendations:

  • Recommendations from product categories like videos and fashion
  • Follow-up purchases of people who did similar purchases to yours.
  • “Inspired by your Wish List”
  • “New for you”

It can be helpful to apply different styling to each type of recommendation. Spotify does that by presenting recommendations as playlists with names and album covers—patterns users are already familiar with:

Sorry, this post’s images got lost.

Spotify’s recommendations could be improved by showing how they are created. More about that in the next section.

When to present only the best result

Lists don’t always work well. In a checkout process, it can be helpful for customers to be reminded that their new camera needs a memory card. Not so much if they’re sent into a whole product comparison cycle with dozens of memory cards, lenses and other accessories.

If picking a wrong item has just a small time cost for the user, you may just as well present the best result and hide the rest for later. That’s what makes TikTok great: open the app and immediately a video starts playing. Not your thing? Swipe to get the next one.

Present the system well

As I needed a whole article to introduce the basics of recommenders, I don’t think it’s desirable to have products explain in detail how their algorithms work. That said, providing some background on how the recommendations are created can make your product more trustworthy.

Explain how data is collected for recommendations

Adding a recommender system can make every user an involuntary target of surveillance. How transparent are you about that? Facebook published quite a bit about how they collect and use data, but still, many users are shocked when they find out about it. Clear communication about the recommender system should be part of your product’s design, if you don’t want to creep users out with jarringly good recommendations.

I recommend communicating about:

  • What data are collected and for how long (you need this to be GDPR-compliant anyway)
  • How the data are relevant for the recommendations
  • How to opt out of data collection

Explain how recommendations are made

When the UI explains about how recommendations are created, users are more likely to trust them. It also makes it easier for them to pick decide what sort of recommendations they want to pick from. In the Spotify example above, I wouldn’t really know why I’d want to listen to the Taste Breakers playlist. Netflix provides a little more information:

Sorry, this post’s images got lost.

Users may also be interested in the metrics used to optimize the algorithm. Are you aiming for engagement or quality experience? For sales or for positive reviews?

Evaluate

You know what Pinterest, Facebook and Twitter have in common? They all have a recommender system’s output as their main UI element: algorithmically created news feeds. My Twitter and Facebook feeds are full of outrage and other extreme emotions, whereas my Pinterest feed is friendly and positive.

Turns out, that’s by design. Despite being billionaires, Pinterest’s Evan Sharp and Ben Silbermann travel the world to meet users where they live. Just so the can learn about what their product’s influence on people’s lives is. Evan Sharp on the Gadget Lab podcast:

The last 3–4 years, we focused quite a lot on trying to make sure Pinterest has a positive outcome for our users, and in particular […] a positive emotional outcome. I’m happy to say that a majority of our users, they leave Pinterest feeling inspired, creative or optimistic about what’s to come. […] A beautiful object that has bad outcomes for people to me, that’s not great design.

As mentioned, recommender systems need and collect a lot of data. For designers that can be a treasure trove of usage data ready for analysis. But even when used in the best statistically sound way, they may not capture actual users’ experiences well. Arguably, Facebook and Twitter have been influenced too much by the quantitative data. Design-driven Pinterest uses both quantitative and qualitative data and manages to avoid some of the other platforms’ problems.

Recommender systems are about personalization, providing choice and prioritization. They’re key to the user experience of many a product. Designers are essential in setting user-centered goals for the algorithms and making sure recommendations are presented for optimal effect.

Next time you’re adding a list to a design, think of how a recommender can improve it!

This is the second part of our series about recommender systems. To learn more, read Recommender system basics.