At the heart of many products in the information age is 1 critical factor: relevance. A misguided recommendation can instantly shatter trust in your product. As a designer, how do you create a system to provide relevant information to your audience in the right place and at the right time? What is an ideal experience? In this talk, we’ll look at data and design, building feedback into your system, and what you need to know about content and machine learning. We’ll explore in depth case studies of designing recommendation systems and the role of “nudges” in changing behaviour and improving outcomes.
CultureAmp is an employee feedback product, which has led to a lot of Recommender Systems.
Recommender Systems predict preferences and show relevant items. Products like Netflix use this to recommend new shows to watch, for example. They are handy but the design territory is often quite uncharted and the UX guidelines aren’t great.
CultureAmp give people preset questions to get started, but then they recommend more. To guide people on which ones to choose they provide context and likely outcomes for each question.
@didoesdigital giving us 3 things to consider when 'Designing recommender systems'
— LadiesThatUX / Melb (@LadiesThatUXMEL) April 12, 2019
Things to think about:
- choosing outcomes
- design considerations
- improving over time
What values are you optimising for? Cultureamp values good people science. What people often think are good questions don’t hold up under the science, but they are of you need to consider how you can guide customers to pick better questions. Other things to consider are time to decision and satisfaction.
Design considerations include things like designing for trust and experience.
To build trust you can use social proof – draw on trusted networks, so you don’t have to prove the recommendations are good; the user’s existing network provides that proof. Using really clear labels helps a lot as well, eg. Netflix is very clear “because you watched X you might like Y”. You don’t have to guess why you are receiving a certain recommendation.
Experiment with how fast or slow your recommendations update and change. Pinterest updates immediately, making use of the Recency Effect. But when done too quickly you can get the ‘instant takeover’ effect. People get upset when you pollute their feed because they looked at something once. Pinterest reveals why things are showing up and lets you tune it.
People love novelty, so use diversity and serendipity. Spotify will give you a bunch of songs in the same general genre (diversity), with some unexpected recommendations mixed in (serendipity).
Data is not neutral. Recommenders depend on data that will usually have inherent bias, and you will have to work to undo that bias. Pinterest’s search results for “CEO” are markedly different from Google.
- Knowledge based – easiest place to get started, you already have data about your own business
- Content-based filtering – automated recommendations linking similar things
- Collaborative filtering – things like a person buying product A and B suggests they may like C
You can also return hybridized results which incorporate more than one technique. CA’s question recommender does this.
Look for positive and negative feedback signals. Choosing a Pinterest pin is a positive signal, unfollowing a topic or manually reporting something is a strong negative signal.
Netflix infers a negative signal if you abandon a show halfway through and never return; vs binge watching. It also listens to active signals, for example when you manually “like” things.
Use analytics to track what people are doing. Diana’s Typey Type project offers recommendations on which lesson to do next; but allows people to take a break as well. But people rarely skipped – they just wanted to keep typing.
Designers are in a position to influence decisions. We can play an important role in the way these systems are designed and how they nudge human behaviour.
If you work on recommenders, please do share your knowledge on it as there is very little out there!