Envision coffee machines that start brewing just when you think it’s a good time for an espresso, office lights that dim when it’s sunny and workers don’t need them, your favorite music app playing a magical tune depending on your mood, or your car suggesting an alternative route when you hit a traffic jam.
Predictability is the essence of a sustainable business model. In a digital world, with millions of users across the globe, prediction definitely has the power to drive the future of interaction. Feeding a historical dataset into a system that uses machine-learning algorithms to predict outcomes makes prediction possible.
What Is Predictive Modeling?
Predictive modeling uses statistical techniques to predict future user behaviors. To understand the intricacy of the design of predictive analytics, you must dive deep and comprehend what a predictive model is. A predictive model uses historical data from various sources. You must first normalize the raw data by cleansing it of anomalies and preprocess it to fit a suitable format that would facilitate analysis. Then, apply a statistical model to the data to draw inferences. Each predictive model comprises various indicators—that is, factors that would likely impact future outcomes—that are called independent variables, or predictor variables.
Applying a predictive-analytics algorithm to UX design does not result in changes to a user interface. Instead, the algorithm presents users with relevant information that they need. Here’s a simple illustration of this capability from the ecommerce domain: A user who has recently purchased an expensive mobile phone would likely need to purchase a cover to protect it from dust and scratches. Therefore, that user would receive a recommendation to buy a cover. The ecommerce site might also suggest other accessories such as headphones, memory cards, or antivirus software.
Here are some other examples of predictive modeling. Spam filters use predictive modeling to identify the probability that a given message is a spam. The first mail-filtering program to use naive Bayes spam filtering was Jason Rennie’s ifile program, which was released in 1996. Bayes theorem predicted which email messages were spam and which were genuine. Facebook uses DeepText, a form of unsupervised machine earning, to interpret the meaning of users’ posts and comments. For example, if someone said, “I like blackberries,” they might mean the fruit or the smartphone. In Customer Relationship Management, predictive modeling targets messaging to those customers who are most likely to make a purchase.
How Predictive Modeling Works
A predictive model forecasts outcomes using data mining and probability. Each model has a number of variables that influence the outcome, but are independent of each other. For example, a student’s test score depends on the number of answers that are correct, which is an independent variable. A dependent variable would be the student’s grade.
A predictive model consists of the training data for a particular type of problem and a learning algorithm that you apply to the data to generate output. To check performance, validate the training data against actual data. At the heart of machine learning are its algorithms. In linear regression, if you try to establish a functional relationship, the linear equation would be [Y = a + bX], where Y is the dependent variable, X is the independent variable, a is the intercept, and b is the slope of a line.
Anticipatory design is a UX design pattern that is gaining in popularity. Its intent is to reduce users’ cognitive load by making decisions on their behalf. Imagine that you want to buy a smartphone, but are bombarded with choices as soon as you land on an online-shopping portal. Having too many choices often results in decision paralysis, a state in which a customer cannot decide. In such a scenario, a predictive engine could build a set of recommendations based on predictor variables, helping the customer make a decision.
Methods of Prediction
There are various predictive models. In this article, I’ll cover decision trees and linear regression.
Use a decision tree to identify the class, or category, to which a target variable belongs. A decision tree is a graph that uses branching to illustrate every possible outcome of a decision.  To explain how a decision tree works, I’ll use the example of a music app’s deciding what song to play next based on certain predictors. A decision tree uses categorical variables for classification. Table 1 shows the training data for this scenario. To keep this example simple, I’ll provide only five predictor variables that influence the decision to play a song.
Table 1—Training data for a music app
A classification algorithm derives from this training data and is applied using the following classification rules.
If time == “morning” && age == “young” && weather == “summer” && emotional state == “Sad” then play == “Melody.”
In our fast-paced marketplace, deciding what new product customers would want presents a major challenge to product-development companies. Creating and launching a new product is a big investment, and there is no guarantee that it will be adopted by consumers. You could use a decision tree to help you decide what new product to develop.
In data mining, regression predicts the value of a target, or numerical variable, by building a model based on one or more predictors. In essence, linear regression shows how variations in an independent variable change the dependent variables. Table 2 shows an example of a linear regression with multiple predictor variables, in which a company is deciding its hiring policy for sales professional and wants to determine whether two independent variables—in this case, a candidate’s years of education and score during the interview process—would predict their ability to generate sales.
Table 2—Predicting a candidate’s ability to generate sales
X1 (Independent variable)
X2 (Independent Variable)
Y (Dependent Variable)
Years of Education
Score in the Interview
Sales generation in a year
Analyzing the correlation between each variable resulted in the following output:
rx1, x2 = 0.9682, rx1, y = 0.8799, rx2, y = 0.7722, in which the correlation for these two variables is 0.935.
The formula for multiple regression is:
Y = a + b1 x1 + b2 x2, in which b1 is 34,356.085, b2 is -3,657.213, and a is 60,049.195.
Now, to predict sales generation for a person with 13 years of education and an interview score of 49, you can apply this formula:
Y = 60,049.195+ (34356.085) X1 + (-3,657.213) X2
Y = 685,881.737
Linear regression is a very powerful statistical technique, and you can use it to generate insights on consumer behavior, forecast sales, and analyze survey data. UX researchers use linear regression extensively in analyzing the results of usability testing.
Leveraging Predictive Models in UX Design
Conversions are one measure of an effective user experience. Using predictive analysis, you can improve conversion rates on an ecommerce site. For example, recommendation systems use data such as consumer interest and buying behavior, then apply a predictive model to generate a list of recommended items.
Amazon uses an item-item collaborative filtering algorithm to recommend products and, thus, can show programming books to a software engineer and baby toys to a new mother. In the future, data will drive the user experience. Patterns that derive from data form the basis of a robust predictive engine that can predict a user’s intent by aggregating several predictors that together influence conversions. Quantum Interface, a startup in Austin Texas, claims to have developed the world’s first and only dynamic, predictive user interface by leveraging natural user interface (NUI) principles. This dynamic, predictive user interface uses the directional vectors speed, time, and angle change to anticipate a user’s intent.
The Importance of Prediction
The disruptive nature of new technologies impacts millions. An unanticipated consequence of such disruptions has been the elevation of user expectations to a level that only artificial intelligence can satisfy. A machine can mimic and even surpass the ability of human intelligence to handle intricate details. Such a system can learn about and interpret users’ past behavior to predict their future behavior. Establishing statistical significance is just one factor. Creating a model that can predict future experience is probably the next strategic consideration to look at in defining an intuitive user experience. Let’s consider a few possibilities.
Data-Driven UX Research
The user-centered design (UCD) process focuses on gaining a deep understanding of users and typically comprises three main phases: discovery, design, and evaluation. The discovery phase starts with identifying the underlying problem, involving key sources such as internal stakeholders, external customers, the Web analytics team, and data from feedback forms. The team analyzes the data they’ve gathered to draw insights from it. However, what teams sometimes fail to do is to validate those insights—possibly because of a time crunch or budget constraints. Thus, the team may fail to realize the potential of statistical models or to apply them to draw the right inferences.
Let’s look an example. Imagine there were a UX redesign effort for an ecommerce portal to resolve a few key user-interface (UI) issues that users had reported: button visibility, text legibility, information findability, a complicated checkout process, and navigation issues. The best way of analyzing these issues would be by creating a problem matrix. Jeff Sauro and James R. Lewis discuss a UI problem matrix in their book Quantifying the User Experience. A problem matrix lists each UI problem, along with user names and descriptions, and a severity rating that takes into account both the observed problem frequency and the problem’s impact on users. From an analytical point of view, a useful approach to identifying UI problems is connecting them with the users who experienced them, as shown in Table 3.
Table 3—UI problem-matrix diagram
4/6 = 0.67
Complicated checkout process
2/6 = 0.33
Using the problem matrix, you will be able to identify patterns in the historical dataset as the problem variables change over time and use them to construct predictive models. For example, they may predict that a particular user segment would always face problems comparing products or struggle with the checkout task flow.
Data-Driven User Personas
A general problem with personas is that they are not based on firsthand customer data or, if they are, that the sample size of the data set is not statistically significant. Creating data-driven user personas is a great way of representing target users. Alan Cooper first advanced the idea of creating such personas, creating these models from data he’d gathered by interviewing seven or eight users. Current practice often differs from best practice. To expedite the process, teams often skip interviewing users and instead create personas based on their intuition and empathy, employing their own mental models and stating problems in terms of their own characteristics or motivations, which may be completely different from those of the target user groups. This process leads to bias.
Other big mistakes with personas result from leveraging information that is not relevant to the context and from using only qualitative data. Agents of change are not representative of the typical mental model of users. Plus, mental models vary among different cultural groups and change depending on the user’s emotional state, life-changing events, the attainment of mental maturity, and the observation of new things. In such cases, applying a statistical model can deliver substantial results. Once you apply a predictive model, you can forecast outcomes even if it’s necessary to add new predictor variables to the existing model.
Data-Driven Decision Making Prior to Interaction Design
Interaction design is the most challenging part of the UX design process. The proliferation of new form factors such as mobile devices, tablets, and wearables—with their multi-touch screens and speech input—have further increased the complexity of designing applications. Dealing with such complexities often detracts from an emphasis on meeting the needs of target user groups.
Let’s look at an example how data analysis fits into the process. Envision a Web application for which the dominant problem that a group of users has reported is with navigation. The application uses a hover menu, forcing users to reach a destination through hover tunnels that become quite annoying and challenging for users, especially those with motor disabilities. This challenge would lend itself to running an A/B test with users to gather statistically significant data that would help to determine a new navigation design. It would probably be necessary to find out what navigation pattern—a navigation hub or a fixed navigation bar—would work better for a smartphone. Over time, test results would demonstrate what navigation patterns would be most desirable for particular form factors, which a UX designer could employ in creating future interaction-design solutions.
Measuring Cognitive Changes
The detection of an inconsistency is a strong trigger for a cognitive change.  A location change will likely lead users to do different things. Exposure to a different culture might prompt them to learn a new way of doing things. Their perceptions and judgment might take a new direction. Every new learning experience has the potential to change their viewpoint and actions. If we can predict and measure such changes and their impacts on decision making, we can establish an anticipatory design pattern to address them.
History can deliver rich insights and opportunities to think ahead. People can learn from their past mistakes and see a better way forward. Predictive models rely on historical data sets; therefore, it is important to record what is occurring in the present to build the right future. With the advancement of machine learning, data science is delivering the next big leap in the digital world.
 Wang, John, editor. Encyclopedia of Business Analytics and Optimization. Hershey, PA: Business Science Reference, 2014.
 Johnson-Laird, P.N. “Mental Models and Cognitive Change.” Journal of Cognitive Psychology, March 20, 2013.
Alipta has been part of the Web IT industry for over 17 years. A seasoned and accomplished UX consultant with rich experience across various domains, he has deep experience in defining and executing digital strategies for B2B and B2C customers. Blending technology and design, he has helped more than 30 global customers in their digital-transformation journeys—focusing on solving everything from core user experience problems to building responsive prototypes. At present, Alipta is focusing on independent research in the areas of Web accessibility and multimodal interactions, building predictive UX models, and mobile user-interface design patterns. Working with the Stanford Crowd Research collective as a UX researcher, he contributed to an initiative to build a sustainable crowdsourcing platform. Alipta completed his Bachelor’s in Visual Arts at the Indian College of Arts & Draftsmanship in Kolkata. He is also a management graduate from IISWBM, in Kolkata. Read More