Designing Adaptive User Interfaces

Abstract English

User interfaces today are mass products developed according to the one-size-fits-all principle. Although the underlying user-centered design process suggests that the focus is on the individual user with their needs and preferences, the focus is actually on the averaged requirements of the entire user group. Especially in the case of interfaces whose target group is very broad and diverse or that are used in very different application contexts, individual needs and requirements of single users arise that often conflict with each other. To balance these out, compromises must be made when designing user interfaces. To better meet the individual needs and preferences of individual users, user interfaces should be personalized. A promising approach to do so is to design adaptive user interfaces. Adaptive user interfaces can detect individual needs and preferences during their use and automatically adapt accordingly to improve the individual’s user experience. This thesis presents the process for designing adaptive user interfaces. For this, characteristics are identified that promise a beneficial application of adaptive user interfaces. The necessary steps to define a feasible adaptation strategy are presented. The automated adaptation process’s technical fundamentals are explained along with the decisions designers must make to ensure its success. Various purposes and approaches to designing adaptive user interface elements are presented along with corresponding design guidelines. The application of the presented methods, processes and design guidelines is demonstrated in an exemplary case-study. In this case study, an existing smartphone widget to control smart light is made adaptive to provide users with a personalized choice of lighting scenes in any situation.

Adaptive user interfaces

The goal of adaptive user interfaces is to best meet the individual‘s or group‘s preferences, needs, and desires by altering aspects of their structure or functionalities (Benyon, 1993). To do that, the user status, the system task, and the current situation during runtime need to be constantly monitored (Rothrock et al., 2002). This way, it is possible to assess the ideal state of the interface at any time and adapt the UI‘s aspects like data representation, the way of interaction, and the assignment of tasks accordingly (Gullà et al., 2018). Thus fast-changing aspects of the user‘s environment can be taken into account in the same way as changes of more persistent user characteristics over a longer period of time.

“Adaptive systems are based on the principle that the system should be capable of identifying those circumstances that necessitate adaptation, and accordingly, select and effect an appropriate course of action.” (Gullà et al., 2015, p. 48)

The adaptivity does not necessarily have to affect the whole UI. Often single adaptive elements within an otherwise static UI already provide a benefit. Thus the spectrum of AUI “can range from an interface which switches fonts to suit the preferences of a user, to an interface which builds and evaluates a model of the user in order to improve the effectiveness of communication between the computer and the user.” (Browne, Totterdell, et al., 1990, p. xi)

The process

The basis of the process to design adaptive user interfaces still is the user-centered design process. Especially because in most examples of adaptive user interfaces, only a part of the user interface is adaptive. Adaptivity adds four major tasks to be performed by an inter- disciplinary team of developers, data scientists, and designers. These tasks are:

Deciding for adaptivity

The demand for UIs that perform well in various usage situations is growing due to trends like mobile computing, digital services accessible on several different devices, and global services that target a broad user group with different backgrounds and life situations. Does this mean we should make every interface adaptive? Not necessarily. The adaptivity comes at a cost. The design and development process of AUIs is more complex than developing a one-size-fits-all mass interface and, therefore, more expensive. Additionally, if not done right, AUI could violate an interface’s usability. Therefore, it is important to analyze the purpose of a UI project and the possible usage situations early on to decide if adding adaptivity would be beneficial and how complex the adaptive system needs to become to add value for the user. In many cases, single adaptive elements within an otherwise static UI are sufficient to achieve the desired result. 

The three main characteristics of user interfaces that each suggests it would benefit from adding adaptivity are:

• Changing contexts of use
• Heterogenous user group
• Changing user needs over time

Changing contexts of use

Adaptive user interfaces can adapt to changing parameters of usage situations. Such parameters are defined as the context of use and reflect the specific conditions under which the user interacts with the application through the user interface (Zimmermann et al., 2014). To increase UX, such conditions should be reflected within the UI design. 

“If a piece of information can be used to characterize the situation of a participant in an interaction, then that information is context.” (Dey & Abowd, 1999, p. 4)

“Context is any information that can be used to characterize the situation of an entity. An entity is a person, place, or object that is considered relevant to the interaction between a user and an application, including the user and applications themselves. […] If a piece of information can be used to characterize the situation of a participant in an interaction, then that information is context.” (Dey & Abowd, 1999, p. 3+4)

To analyze a usage situation for changing context parameters that an adaptive user interface could adapt to the following checklist can be used:

PDF

Heterogenous user group

Humans differ in many aspects. Aspects like demographics, physical characteristics, abilities, knowledge, motivation, personality, and preferences. Other than the volatile and situative aspects of persons in focus when analyzing the context of use for changes, these human characteristics persist over time but differ between single users. Depending on the intended use of an application or a service, the single aspects of these persistent human characteristics become relevant or not. If they influence the requirements that have to be met to accommodate optimal use, they should be considered when designing the interface.

To find out whether the use of an AUI is worthwhile to meet the diverse needs of heterogeneous user groups, a three-step analysis of the intended project is necessary:

An example of improving an interface’s usability by adapting to the individual user’s physical characteristics is the size and positioning of intertable elements on smartphone displays. Smartphones are often held and controlled with one hand resulting in the thumb being the only finger to initiate touch events on the screen. Today’s smartphones’ screen sizes are often too large to reach all areas with the thumb. Some smartphone apps take this into account by avoiding placing interactable elements in the upper third of the screen. This optimizes the apps for one-hand usage. However, different thumb sizes of users are rarely taken into account. Assuming that large thumbs have a greater range, but small thumbs have greater precision in hitting smaller elements, the individual user’s thumb-size could be assessed to resize the interactive area accordingly.

Changing user needs over time

User needs are the primary source to derive requirements for UI-Design. In user-centered design, everything revolves around user needs. User needs and the resulting requirements of users towards interfaces cannot be seen as static. Instead, they are subject to time and task-dependent individual differences that need to be considered (Haaks, 1991). They are also dependent on both the context of use and the user’s human characteristics, so user needs appear in all three main categories that outline the beneficial use of adaptive user interfaces. Next to situationally conditioned needs due to contextual changes and the differences in user needs between single users, changing user needs of single users over a longer period of time with several single usage situations can also be an indicator for the beneficial application of adaptive user interfaces.

Additional indicators

Additional indicators for the beneficial application of adaptive user interfaces are:

• Multiuser interfaces
• Series of repetitive steps
• Many options
• Limited output modalities
• Limited explicit input modalities

Defining and validating an adaptation strategy

The design of adaptive user interfaces follows the iterative user-centered design process with its phases of research, concept, design, develop and test. While it often is sufficient to follow these phases one after another in static user interface projects, in adaptive user interface projects, the individual phases must be more interrelated. Since hardly any standards for the development of adaptive user interfaces have been established so far, it is important to test single aspects of the envisioned adaptive system as early as possible for technical feasibility and user acceptance by involving the necessary disciplines early on. The extra effort necessary to develop AUIs compared to static UIs only pays off when the user benefits from the adaptation. The adaptivity needs to serve the user, not the system. To do so, the prediction of beneficial adaptions needs to reach a certain level of accuracy. Improvements in the accuracy of adaptive user interfaces have a strong effect on performance, utilization, and user satisfaction with the interface (Gajos et al., 2008). Therefore, it is important to understand all important aspects of adaptive user interfaces early on and validate their feasibility before investing effort into something that cannot be implemented sufficiently. To avoid this, the research and concept phase of adaptive user interface projects include 4 important steps:

© Florian Schröder

Selection and processing of suitable data sources

In a static user interface, defined user input results in predefined output. For an adaptive user interface, the output is not predefined but determined by many variables besides the explicit user input. A common goal of adaptive user interfaces is to reduce user effort by evaluating information about the user and the usage situation to simplify the interaction of the user with the interface. To assess a situation correctly, determine user tasks and goals, and even predict user behavior, lots of information are necessary. Requiring the user to input all this information actively would cancel out the adaptive user interface's positive effect in terms of efficiency. It would turn the adaptive user interface into an adaptable user interface. That is why implicit ways to gather the necessary information without bothering the user have to be found.

To perform fitting adaptations for the user, an adaptive system needs to process data in three stages (Weibelzahl, 2002):

• Data acquisition  
• Information inference
• Adaptation decision

Data acquisition

The automated adaptation process of adaptive user interfaces is based on data that is generated and collected during use. The data is obtained through various sources, which can be classified into four categories:

• hardware sensors
• user-behavior trackers
• status trackers
• (online) databases

Information inference

Within the next stage of data processing, an intelligent mechanism infers abstract user characteristics from the raw data. In contrast to the data acquisition stage, the properties are inferred from the input data and not accessed directly (Weibelzahl, 2002).

The inference is based on AI techniques. AI, in this case, is defined as the simulation of human intelligence processed by machines. Standard approaches include Bayesian Networks, machine learning, Case-Based Reasoning, rule-based inferences, and combinations of these (Weibelzahl, 2002).  

The process of information inference with such methods contains two major steps:

1. inference of high-level information
2. inference of user needs and tasks and the resulting requirements towards the UI

To make valid statements about user needs, high-level information is needed that often cannot be measured directly. Therefore in a first step, such information needs to be inferred by processing the measured values with suitable methods. Many of these methods require AI to turn data into actionable information based on pre-trained models.

Inferring user characteristics like needs and preferences from the acquired data and aggregated high-level information is done in a second step of the information inference stage of adaptive systems. Adaptive user interfaces records and analyzes the user‘s actions and interactions, together with the prevailing context conditions, to learn how the user acts within certain conditions, what the user prefers to choose from existing options in a particular situation, or how they react to certain information.

Adaptation decision

In the adaptation decision stage, the system decides about concrete adaptation steps based on the inferred abstract user properties (Weibelzahl, 2002). The inferred abstract user characteristics trigger the adaptation and function as input variables to decide what actually gets adapted. This process defines the adaptive behavior of a concrete situation to meet the overall adaptation goal. The fitting adaptation decision is made with respect to the adaptable parameters of the AUI and defined adaptation rules.

Designing adaptive user interfaces

Once the inference mechanism calculates necessary adaptations, they get applied to the adaptable parameters of the adaptive user interface following the adaptation rules. Defining these is the responsibility of UI/UX Designers. When designing the user interface , they define the spectrum of possible adaptations and the framing parameters in which the adaptive system can act autonomously after shipping the user interface to the users. Therefore, designers have to foresee the spectrum of beneficial adaptation and provide the necessary flexibility within the user interface design. The biggest challenge is to reconcile the sometimes conflicting requirements of allowing users to benefit from personalization in the best possible way while at the same time not violating the usability of the user interface. Different approaches to adapt user interface elements and purposes for adaptations are possible

In general, all aspects of user interface elements can be adapted. These aspects include, for example, user interface control structures like menus and toolbars, content presentation within user interfaces, and dialogues with the user. An interactive tool was created that list common purposes of adaptive GUI’s that can be supported by different approaches of adaptations which in turn can result in changing different attributes of user interface elements. The single presented expressions are not always entirely separable. Examples can often be assigned to more than one presented approach and purpose. The connections drawn between attributes, approaches, and purposes are partly derived from research and partly result from subjective assessment without making any claim to completeness. They are intended to provide an overview of the possible uses of adaptivity in user interface design and to inspire designers to apply adaptivity in their user interface solutions.

© Florian Schröder

Design guidelines

1. Apply adaptivity to suitable use cases only
2. Make it attractive to use the adaptive part 
3. Allow users to use the UI the way they know it 
4. Set expectations right
5. Make adaptations comprehensible
6. Be transparent about the quality of recommendations
7. Design for privacy and trust
8. Design for accuracy and reliability
9. Allow for user correction
10. Create good presets Design against the experience bubble

Case study

Problem

Jim, a 30 years old guy living on his own in a 3-room apartment, was using smart light bulbs in his living room for one year before equipping his whole flat with them. Those light bulbs can be controlled via a smartphone application. Depending on the bulb type, they are dimmable and can change their color within the white spectrum or the whole color spectrum. One of the main benefits of smart home appliances like connected light bulbs is that single devices do not need to be controlled individually anymore, each with their own respective switches, but can be controlled with one single action (Ringbauer & Heidmann, 2006), e.g., by selecting a so-called scene via a smartphone application or remote control. Within such a scene, the settings for a group of light bulbs are defined, and with only one tap, it is possible to activate them. Jim created different scenes for his living room for different situations like sports, working, relaxing, or watching a movie. To access his favorite scenes quickly, he uses the widget his smart-lights app provides (figure 16). This widget allows him to access five predefined scenes or single functions directly from his smartphone‘s home screen without opening the app. The widget is very adaptable. He can adapt and create scenes for his needs and choose his favorite ones to be displayed. He was happy with this function for a while until he started to create more and more scenes for recurring situations and equipped his whole flat with smart bulbs. The five options he could display in his widget were not enough anymore, not even for his living room only. He had to use the app again. To set the fitting scene, he has to unlock his phone, open the app, select a room, go to scenes and select the scene. That is three interaction steps more than when selecting a scene from his favorites widget. Setting his lights became less efficient the more scenes he created and the more connected light bulbs he integrated into his home.

To solve Jim‘s problem, it was analyzed that making it adaptive could be a solution. This was done from the point of view of the company selling smart lights and offering the app:

The company‘s goal is to offer an efficient way for its customers to set multiple lights within the home. They hope that users will buy more light bulbs when the effort controlling them does not increase with the number of light bulbs installed. They want to offer a solution to control the lights via the smartphone, as such interface devices are widespread. They do not want to limit their target group as basically anyone should buy and use their smart light bulbs.

Solution

The analysis revealed that making the smart lights widget adaptive is promising as the following characteristics apply:

• Changing context of use (see filled out checklist)
• Many options
• Limited output modalities

User research revealed that the most influential parameters in the decision how to set the lights are the users room and their time-based routines (see ratings in gallery below).

Based on the research results an adaptation strategy was formulated and metrics were defined (see in gallery below). 

The feasibility analysis resulted in the possibility to make the necessary data sources available to an adaptive system that infers probabilities for each light scene available to be picked by the user. 

The light widget was redesigned to become adaptive. The room the user is located in has the biggest influence on the inference of probable scenes to be chosen by the user. In most cases, the entire set of displayed scenes is replaced when the user changes rooms. To make adaptations comprehensible (guideline 5) the adaptive widget should reflect this within its design. This is done by changing the background color of the whole widget depending on the room the smartphone is located in. Additionally, the room’s name and icon are displayed. This allows the user to see at a glance whether the assessment of the current room is correct or not, without having to apply undue additional cognitive load. The user changing rooms is reflected by a vertical 3D animation of the whole widget content. Additionally, the user can undo the adaptation with a vertical swipe on the widget to allow corrections (guideline 9). 

The inference mechanism calculates the probabilities of all scenes independent of the room they belong to. This probability distribution gets reflected by the order of scenes displayed in the adaptive widget with decreasing probability from left to right. For newly created scenes by the user and those that are not used regularly, the system lacks the experience to properly calculate their probability to be picked in a given situation. Thus they randomly get presented on the fifth position once the room they belong to is identified. This allows refinement of the user profile by letting users discover scenes without additional effort (guideline 11). Additionally, users can reach scenes beyond the five presented once by a horizontal swipe on the scenes to allow further discovery and correction (guideline 9). This possibility is made visually clear by partly displaying the 6th scene in the sorted series. If the probabilities of scenes to be picked and thus the order of scenes presented changes while the widget is visible, this gets supported by horizontal animations (guideline 5). Left moving scene(s) which probability increased move in front of the right moving scene(s) which probability decreased towards their new position. 

Users are able to activate the widgets personalization function via the widget setup menu or switch back to the original customized version if they like (figure 51). A text explains the adaptive function of the personalized version (guideline 4) and asks the user to activate the room detection and the behavior tracking (guideline 2+7). Once activated a dashboard displays the inferred probabilities of the single scenes and the probabilities of which room the user currently is in. Thus it is comprehensible and transparent which data the order of scenes is based on (guideline 5+6). A button offers the user to calibrate the room detection to improve accuracy (guideline 8). If the room detection is not precise enough, the user will recognize this via the dashboard and thus is motivated to go through the calibration process. If the room detection works well without, the extra user effort can be saved. If the system detects the need for a calibration itself, e.g. due to many user corrections, the user is requested to perform the calibration via push notification. Scenes that were already part of the user-customized widget should be preferred in the beginning when the collected data does not yet provide enough insights into the user’s habits. Combined with only displaying scenes from the detected user’s room already provides a benefit to users. Thus a preset can be generated from the available data (guideline 10).

© Florian Schröder

Download Thesis:

PDF