A second option to refine adaptation rules is a form of dy-namic knowledge acquisition, which observes the user’s be-havior and infers mistakes and/or improvement possibilities.
For example, if a user moves the mouse pointer for a cer-tain time without clicking, this could indicate that he or she is looking for something, and hence, needs more support. So far, work on inferring user interface adaptations from mouse movements or other user input has proven feasible as a way to classify users into novice or expert [Hurst et al., 2007], or de-rive user interface constraints for people with motor impair-ments [Gajos et al., 2008]. Previous studies also demonstrate that one could draw conclusions on certain cultural dimen-sions from a user’s mouse movements and navigation behav-ior [Kralisch, 2005, Heimg¨artner, 2005]; however, these ob-servations did not point towards the possibility of inferring on user interface preferences as envisioned by our adaptation rules.
The prerequisite for applying user interaction tracking is Inference rules
the definition of inference rules, which map certain interaction
patterns to adaptations of the user interface. Thus, we have to determine (1) which parts of the user’s interaction with the application are feasible and useful to record, (2) what we can conclude from certain interaction patterns (i.e. a classifica-tion of users into certain categories), and (3) what informaclassifica-tion we can derive from interaction patterns that can be stored in CUMO for use by other applications.
As to the feasibility, Schmidt et al. [2007] have shown an approach to semantically annotate and extend a Javascript/
AJAX-based web application with an interaction tracking com-ponent, which records mouse movements, clicks, or keyboard input. Previous work has also derived different interaction statistics for classifying users into different learner proficien-cies with the help of eye tracking [Amershi and Conati, 2007], skill levels [Hurst et al., 2007], or derive special needs, such as the reading speed for people with disabilities [Stephanidis et al., 1997]. So far, however, it is unknown whether user in-terface interactions could also reveal other user characteristics and preferences.
Theoretically, the inference part of the user modeling pro-cess can be handled independently of previous information
Discussion of in-teraction statistics
and their meaning contained in the user model, in our case of culture. Thus, ob-servations of the user interactions and inferences on further adaptations are not necessarily restricted to the user’s cultural background, but can include an upper level of observation:
Is the user able to cope with the adaptations? Are there be-havioral restraints that might point to a need for correcting those initial adaptations? As mentioned earlier, mouse hover-ing could indicate the need for more support, but it can also mean that a person is simply reading. Coping well with a user interface is usually characterized by determined mouse movements (forming a straight line at a certain speed), and few errors (e.g. the use of the browser’s back button, or open-ing and closopen-ing a dialog without fillopen-ing it out could indicate errors). Hence, these upper level observations mainly hint at
4.4 Refinement of the Adaptation Rules with Machine Learning 79
the skill level, which in our case could complement informa-tion about the computer literacy in the cultural user model.
In addition, we were curious to find inference possibili- Inferring information about user model aspects other than residences
ties to gain information about other parts of the cultural user model (i.e. knowledge that has not been explicitly acquired in the initial registration process). Specifically, the question of whether it is feasible to use the interaction tracking to gain or verify information on the user’s language/reading direction, education level and form of education, political orientation, or religion. While the level of computer literacy can be extracted from interaction statistics, these other aspects in CUMO do not necessarily express themselves in the user’s interaction with a user interface. In particular, it seems to be more real-istic to estimate someone’s political orientation, religion, and the form of education from information about former and cur-rent residencies than via deduction from mouse movements.
Computer literacy, in fact, often correlates with the education level (see discussion in 3.2). Our approach, therefore, at first relies on deriving the computer literacy only, and leaves the verification of automatically derived assumptions about the education level to the user.
In addition, the reading direction can be inferred from the user’s language (assuming that we can deduce this informa-tion from the country of current and former residence). Al-though there is the possibility to infer the reading direction from the unconscious focus of the eyes with the help of eye tracking [R ¨ose, 2005, Amershi and Conati, 2007], we have re-frained from this due to the fact that eye tracking has not yet been proven to be an everyday solution.
From this analysis of previously employed interaction sta-tistics and our own thoughts on possible inferences from the user’s interaction, we have compiled a list of interaction tics. Table 4.4 gives an overview of those interaction statis-tics that are applicable to our approach, and briefly describes the interpretation of them into inference rules. Note that this
interpretation could possibly also effect the size of buttons, and the distance between user interface elements for older or motor-impaired users, as suggested in [Gajos et al., 2008]. For now, however, we only make inferences on culture-related changes to the interface according to the adaptation rules in Table 4.3. As suggested in [Hurst et al., 2007], our inference rules aim to classify users without the previous establishment of a task model (i.e. the interaction sequences are unknown).
The table reveals that the user’s interaction with the inter-face mainly provides hints about the user’s familiarity with the software, or the computer literacy in general. Thus, we
Overview of
in-ferable aspects can derive the need for a different support level, non-linear or linear navigation, the number of functionalities, and different content structuring. These aspects are predominantly affected by the dimensions Power Distance, and Uncertainty Avoid-ance (cf. Table 4.1 and 4.3), and could therefore complement or correct the information gained in the initial acquisition pro-cess.
The implementation details of our approach to user inter-action tracking are described in 5.7, where we also present a short evaluation of the classification system of users into low and high computer literacy.