Talk: “Designing Recommender Systems” at Web Directions Design 2019
I spoke at Web Directions Design 2019 at the Arts Centre Pavilion in Melbourne, exploring the user experience design side of recommender systems for predicting user preferences about items and showing relevant items.
Web Directions delivered yet another excellent conference. 3 days of wonderful talks about design including ethics, scaling design, and the practice of design. Check out the #design19 hashtag on Twitter.
About the talk
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.
Here are the slides:
The big ideas
Recommender systems predict user preferences about items and show relevant items.
To design a successful recommender, you need to establish your desired outcomes, which might be:
- Business impact: such as profit or customer lifetime value.
- Values: for example, good people science.
- Time to decision: try reduce the time it takes people to make decisions in your product to reach value faster.
- Satisfaction: keep people happy, especially in subscription products.
Once you know what success looks like, there are many design considerations, especially around trust and experience. When designing for trust, consider:
- Social proof: surface recommendations from trusted networks (friends and experts).
- Labels & headings: “Recommended for you”, “Because you watched X”, “Created with Y” versus “Sponsored” or “Promoted”.
- Recency and stability: update recommendations quickly in response to users’ changing interests, but maintain a stable experience.
- Transparency: give insight into the black box of recommendations e.g. “This pin is inspired by your board Hamburgers”.
When designing for experience, consider:
- Diversity and serendipity: instead of 17 remixes of one song, recommend a diverse variety of songs in a mood or genre, and facilitate happy accidents with surprising recommendations.
- Visual clarity: experiment with how you present content, even without changing the recommendation itself.
- Data is not neutral: consider how to address data that is not neutral in your platform.
For an example of this last point, compare image search results for “CEO” on Google Images versus Pinterest.
Three popular recommender techniques include:
- Knowledge based: you know how your problem solves user needs e.g. a mortgage broker can suggest a loan using industry knowledge and understanding of your constraints.
- Content filtering: we can recommend this other book in the same genre using only data known about the items, not the users.
- Collaborative filtering: when people buy X, they also buy Y.
“Hybrid recommender systems” use combinations of recommender techniques, which come with their own hybridization techniques.
To improve predictions, many recommender systems look for positive and negative feedback signals as well as implicit and explicit feedback signals, which you can design into your interface and your recommender system.
Without machine learning, you can inspect data and perform user research to understand which recommendations are successful and make manual adjustments.
If you have ideas about the interface design or user experience design of recommender systems, please share your thoughts! The industry is sore lacking resources on these areas in plain English for designers and researchers.
Links and resources
- Naren Katakam wrote about How Can We ‘Design’ An Intelligent Recommendation Engine? on Medium, covering high-level systemic design of recommendation engines and grassroots-level interface design.
- Nielsen Norman Group studied Individualized Recommendations: Users’ Expectations & Assumptions and discovered that “users appreciate personalized content suggestions and are willing to give up some of their privacy for quality recommendations, while accepting some inaccurate recommendations.”
- Netflix research shares a lot of insights into the development of recommender systems, although in many cases it gets rather technical.
- The Wikipedia page on recommender systems is a reasonably accessible place to get the lay of the land.