AssistU - a framework for user interaction forensics

Users make expensive mistakes, most of which are preventable. In the medical domain for example, this is true for up to 80% of operating errors [Wie08]. These operating errors could be prevented, because they are not the user s fault. Rather, the user interface did not meet the user s personal requirements. This can be adjusted offline, but these requirements might change. Especially when considering usage over a longer period of time (or under different circumstances), the user skills and abilities will change. The main difficulty is to detect a situation during run-time that very likely will lead to an operating error and to determine and execute adequate measures to prevent such an operating error. Especially in industry (e.g., production industry or medicine), costs are high because human lives are at stake. Another difficulty is the systematic integration of a solution into the process of user interface development without the addition of unnecessary complexity. To conquer these problems, this thesis introduces an automated user interaction feedback loop (collect, analyze, decide, act) the AssistU approach. The underlying assumption is that the user reveals more information about himself than merely the activated function. Simple metrics (collect can be measured online. Examples are point of interaction, point in time, or task order. Complex metrics (analyze) can be computed from these simple metrics. Examples are precision (ability to cope with the dimension of an interaction object) or handedness (left-/right-handedness influences the optimal layout). The result is a User Performance Model. The AssistU approach is extending a state-of-the-art model-based user interface generation process the SmartMote model framework. Based on the SmartMote rendering process, adaptation rules (decide) change the presentation model. According to the presentation model, the renderer automatically changes the appearance of the user interface to be displayed (act). This happens after every dialog chance. The feasibility of the approach was shown based on an example user interface generated for the SmartFactoryKL. In a further evaluation, the existence of interaction forensic patterns was shown, which is necessary to allow analyzing the metrics. The following are two of the results of the experiment. In the setting of the experiment, the users only used a small part of the available interaction area (7.5%). Thus, interaction is not random and can be used to predict a loss of precision. Right-handed users could be determined by the location of the centroid of the used interaction area as well as by the difference in the precision of the interaction depending on the location of the interaction object.