In this edition of Ask UXmatters, our expert panel considers how best to influence the design of personalization logic so it effectively delivers what users need. Our panelists discuss recommendation systems, some technologies that support personalization, the difference between customization and personalization, the different types of personalization, and the process for designing and testing personalization.
Our experts explore, in some depth, the various stages of development during which UX researchers and designers can have the greatest impact on the design of personalization logic and the contributions they can make to the process. Finally, we consider the importance of iterative testing and design in designing personalization software.
In my monthly column Ask UXmatters, our panel of UX experts answers our readers’ questions about a broad range of user experience matters. To get answers to your own questions about UX strategy, design, user research, or any other topic of interest to UX professionals in an upcoming edition of Ask UXmatters, please send your questions to: [email protected].
The following experts have contributed answers to this edition of Ask UXmatters:
Pabini Gabriel-Petit—Principal Consultant at Strategic UX; Founding Director, Interaction Design Association (IxDA); UXmatters Publisher, Editor in Chief, and columnist
Jordan Julien—Founder of Hostile Sheep Research & Design
Gavin Lew—Managing Director at Bold Insight
Janet Six—Product Manager at Tom Sawyer Software; UXmatters Managing Editor and columnist
Q: When designing applications that use personalization logic—such as recommendation systems for ecommerce sites—how can one best influence the design of the personalization logic so its output matches the user’s needs?—from a UXmatters reader
“There are a few ways to approach the user experience of algorithms such as personalization algorithms,” replies Jordan. “My favorites use flow diagrams or taxonomies. For something like a recommendation system, I’d start by defining a taxonomy. Tag every product with one or more terms from the taxonomy. This establishes the initial relationships between content elements or products. For example, the Gap might tag a T-shirt as a V-neck, then use that tag to recommend other products they’ve tagged as V-necks.
“But this is only a starting point. I’d also recommend using some form of WEM (Web Experience Management) or XM (Experience Management) software that can learn from users’ behavior. You can serve up unique recommendations to individual users based on how they browse a Web site. Ultimately, the more people use the system, the better the system becomes at predicting what products it should recommend.”
Content Management System (CMS)—These systems help companies organize their content, images, and data and facilitate the publication of content online, but do not support personalization.
Web Experience Management (WEM)—Such systems enable organizations ‘to share content, data, logic, and other elements across channels. ‘[They] introduced rule-based personalization to the online experience and gave the ability to collect user behavior, define personas, and create and provide unique content to the targeted audience.’
Digital Experience Platform (DXP)—According to Gartner, a DXP is ‘an integrated set of technologies, based on a common platform, that provides a broad range of audiences with consistent, secure, and personalized access to information and applications across many digital touchpoints. Organizations use DXPs to build, deploy, and continually improve Web sites, portals, mobile, and other digital experiences.’
“DXPs are gaining preeminence as companies move from simple Web experiences to multichannel digital experiences and personalization evolves from simple, rule-based systems to contextual personalization that is based on machine learning. A DXP leverages everything an organization knows about a user’s interactions with the system across all channels to deliver a personalized user experience.”
Influencing the Design of Personalization Logic
“Personalization requires a deep understanding of users’ needs—whether the role-based needs of well-understood user personas or the needs of individual users for which the system dynamically builds user profiles,” answers Pabini. “Thus, Nielsen Norman Group defines two types of personalization:
role-based personalization—In this type of personalization, user personas have specific, well-defined characteristics based on a role.
individualized personalization—In this type of personalization, also known as one-to-one personalization, a system infers individual users’ characteristics from a history of their interactions—as well as any data they’ve provided—and creates a model, or profile, of each user.
“You should first conduct the generative user research and analysis that is necessary to understand users’ needs. Gathering data about users’ interactions tells us what they do, but not why. Correlating user data with user-research findings gives you a deeper understanding of users’ motivations and needs. As a result, you can create user personas that model typical classes of users, as well as identify the salient user characteristics that a generated user profile should comprise.
“IBM have outlined an eight-step process for planning and implementing personalization. Armed with the findings of your user research, you can have the greatest influence on the design of the personalization logic during the following steps of this process:
Step 1: Identifying the business goals for a personalized user experience—You can help define the business goals by asking the following questions: ‘Who will the users be? What do you know about the users and in what form does this data exist? What information, or content, do you want to provide in a personalized way? How will you measure success?’
Step 2: Developing user and content models—A user model consists of specific attributes of users. Such models may comprehend user personas and/or a model for automatically generated user profiles. This is the most user centered part of the process, and you can bring your understanding of users to bear. A content model determines what content should be personalized for specific users, defines the metadata for personalized content, and determines how you’ll decompose content into chunks. Defining metadata requires knowing what attributes of the content are meaningful to specific users. For example, to facilitate the mapping of users to content, the user model might include the property topics_of_interest, while the content model would include the corresponding property topics.
Step 3: Choosing a personalization approach—that is, rule-based and/or recommendation-based personalization. Make sure the personalization approach your team chooses maximally supports users’ needs. Machine learning enables systems to perform tasks by generalizing from examples, delivering outcomes that improve over time. For rule-based personalization, you’ll define business logic based on the user data you’ve collected and the existing content. Recommendation-based personalization employs a recommendation engine that uses a collaborative-filtering process in which a site collects data about users’ interactions, then uses that historical data to determine what content to deliver to specific users. Predictive segments determine what content to display during a user’s current session based on the types of content that were favorably received by clusters of users with a history of similar behaviors. Thus, collaborative filtering does not necessarily depend on knowing users’ attributes.
Step 6: Creating personalization rules—These rules define the business logic behind personalizing the system for a specific user or class of users. Those of us who are not math wizards can define such rules in pseudocode, in the form of simple conditional expressions such as: ‘If the user is a registered user, then display the user’s favorites.’ You could also define personalization rules as decision trees or flowcharts.
Step 7: Testing the personalization rules—As for any software, usability testing and analytics can determine how successful the software is in delivering the personalization a user needs.”
Personalization Versus Customization
“The first thing to note is that a very small percentage of the population personalizes software themselves,” says Gavin. “Really, when was the last time anyone went into settings?”
“True, but Gavin is confusing—or conflating—personalization and customization, as many people do.” remarks Pabini. “Customization refers to the changes users make to an operating system or application by modifying settings—for example, in iOS, selecting a background image, or wallpaper, for the lock screen or home screen. Customization requires considerable effort on the part of the user. In contrast, personalization is something applications do on behalf of users—tailoring a user experience, workflow, navigation, search results, functionality, content, or notifications to their unique needs by leveraging what the system knows about them. It requires no effort on the part of the user. Marketing and sales are at the core of many implementations of personalization, including many recommender systems. However, personalization can efficiently deliver exactly what a specific user needs in any type of application, helping the user to be more effective in completing tasks.”
Taking a Proactive Approach
“What do people do when they encounter recommendations?” asks Gavin. “Most people overlook them or roll their eyes and don’t do anything. So we need our designs to be smarter. Thus, the recognition of cues is very important. Consider the recommendations section in Google News—following the headlines and other sections—which presents a list of articles. This content is personalized. While you may not recognize this, an algorithm has pulled links to articles for you.
“The cues that can support better experiences are now available, but you have to collect and use the data. One needs to understand whether people stare—and for how long—or just whiz right by. When people click, how should you weight that? Can you measure the length of time someone spends reading—before returning to a page of links—to understand the value of the article to the user? Did the user actually read the article or click the link accidentally? Such cues can tell you whether you are hitting the mark and offering a benefit. Are clicks increasing or just random? What is people’s threshold of attention relative to other sections? This can tell you whether the design is doing its job. These are the sorts of questions that help show the value of personalization.
“If you do all of this well and understand how people will use a design, this allows you to take a proactive approach,” continues Gavin. “For example, Google News asks whether you are still interested in X and, with a press of a button, you can say Yes. But you must already have earned users’ trust in your personalization to allow such a proactive approach.”
“Ultimately, you have to understand the context. Algorithms are going to come up with recommendations, but the question is whether they are right. What are people trying to do and when are you presenting recommendations? However, these are still just your best guess. You need to understand and look for cues that tell you whether you’ve hit the mark.”
Testing and Iterating Our Designs
In addition to creating good UX designs that consider the needs of both the users and those who analyze the business success of personalization, we must test our designs in users’ real environments. For a rules-based personalization system, we must also develop software within a rapid, iterative environment that lets us frequently update the personalization-logic engine so it continually becomes better.
While personalization seems like a smart thing to do, if we do not ensure that our designs are effective, personalization logic may deliver no better results than the typical salesperson who tries to get the attention of passersby with phrases such as: “How are you today?” “Would you like to save money?” or “I have a special deal for you!”
The personalization logic needs to provide recommendations that are genuinely of interest to the customer. Otherwise, the customer will ignore them. The creation of a poor personalization-logic engine then becomes a very expensive way of creating a new form of banner blindness in the user.
Dr. Janet M. Six helps companies design easier-to-use products within their financial, time, and technical constraints. For her research in information visualization, Janet was awarded the University of Texas at Dallas Jonsson School of Engineering Computer Science Dissertation of the Year Award. She was also awarded the prestigious IEEE Dallas Section 2003 Outstanding Young Engineer Award. Her work has appeared in the Journal of Graph Algorithms and Applications and the Kluwer International Series in Engineering and Computer Science. The proceedings of conferences on Graph Drawing, Information Visualization, and Algorithm Engineering and Experiments have also included the results of her research. Read More