Running head: VISUALIZING INDIVIDUAL DIFFERENCES IN WEB NAVIGATION Visualizing Individual Differences in Web Navigation: STRATDYN, a Tool for Analyzing Navigation Patterns
چکیده
Navigational behavior on the Web can be analyzed with different methods. Log file data are an important source of behavioral traces of navigation. In this paper, we first discuss existing approaches to the classification and visualization of movement sequences which are important for understanding Web navigation. We then present STRATDYN, a tool which provides meaningful quantitative and qualitative measures from server-generated log files as well as easy-to-follow visualizations of navigational paths of individual users. We demonstrate the usefulness of this new approach by reporting results of two studies (with 44 students in education and vocational training) which show that navigational effectiveness is positively related to the ability to concentrate and selectively focus attention as measured by the D2 Test of Attention and the FWIT, a German version of the Stroop Test. Finally, we discuss implications for further research in this area and for the continuing development of the presented approach. Visualizing Individual Differences 3 Visualizing Individual Differences in Web Navigation: STRATDYN, a Tool for Analyzing Navigation Patterns The question of how people navigate through hyperspace, what decisions they make, and why they make these decisions is important for those concerned with computer-mediated learning as well as for interface designers and authors of hypermedia information spaces. Although users are nowadays more familiar with hypertext, and the quality of hypermedia design continues to improve, users still complain about "being lost in hyperspace" and suffer from cognitive overload (Edwards & Hardman, 1989; Wright, 1991). The questions as to why users may have problems when navigating in complex information spaces can be addressed with a broad spectrum of research methods, such as think-aloud protocols, interviews or questionnaires (Chen & Rada, 1996; MacGregor, 1999). Due to the inherent shortcomings of self-report data, it is however desirable to obtain "objective" behavioral data regarding navigational paths and user actions, which can be provided by server log files. However, obtaining, preparing, analyzing, and interpreting such data is not easy. Getting a full recording of participants’ Web interactions is the first hurdle researchers have to take (Ellis, Jankowski, Jasper, & Tharuvai, 1998). The biggest challenge, however, usually lies in analyzing and interpreting log file data. While methodological procedures have been developed for visualizing and analyzing large-scale log file data for data mining purposes (e.g., Spiliopoulou, 1999), few effective procedures exist for the psychological Visualizing Individual Differences 4 researcher interested in the navigational behavior of individuals in response to different task demands or conditions (Unz & Hesse, 1999). In this paper, we present a new approach to the classification and visualization of navigation behavior which provides the researcher with quantitative measures as well as with an easy-to-follow graphical representation for detailed qualitative analysis. In the following, we first discuss existing approaches to the classification and visualization of movement sequences which are important for understanding Web navigation. Second, we discuss requirements for the visualization of individual navigation graphs and then present the approach taken with the tool STRATDYN. Third, we show that individual differences in the ability to selectively focus attention when navigating the Web can be analyzed with the help of the qualitative and quantitative data generated by STRATDYN. Finally, implications for further research in this area and the continuing development of the presented approach are discussed. Formalizations of Navigation Behavior Navigation behavior is often formalized in terms of measures. A measure is a one-dimensional variable computed from the recorded navigation history, in our case, Web server logs. Typical measures include the total number of pages visited to solve a task (e.g., Kim, 2000; Wandke & Hurtienne, 1998), the total time needed to solve a task, or average times spent on single pages (e.g., Hall, Balestra, & Davis, 2000), or to what extent pages supposed to be accessed together were in Visualizing Individual Differences 5 fact accessed together (e.g., Hall et al., 2000). However, such measures treat navigation histories as essentially linear entities, neglecting the non-linear structure of hypertext and navigation sequences, where re-visits to nodes are common. The modeling of navigation as a graph preserves this information. A Web site, or another set of hyperlinked pages, itself can be modeled as a graph (or nodeand-link structure, cf. Conklin, 1987). Navigation traverses this graph, visiting some nodes but not others, revisiting some nodes, but considering other nodes only once each in long linear sequences. In this traversal, patterns may exist. Patterns can be particular sequences of nodes (e.g., [home page, general-infopage]), or more complex regularities in behavior. For example, one may look for ”routes through the data that do not cross any node twice” as opposed to ”routes through the data which return to the node where they started”. The occurrence of such patterns can be counted, giving rise to indices like ”pathiness” or ”ringiness” used to denote the number of such routes by Canter, Rivers and Storrs (1985). So if patterns can be rigorously specified, they define measures on the data. Here, we model navigation behavior in terms of navigation graphs. A navigation graph is a directed graph consisting of one node per distinct URL visited and one edge (link) per movement from one URL to the next. Each edge is labeled by the number of the step at which this movement is made. Visualizing Individual Differences 6 Navigation and Search Formalization as a graph is related to modeling in the problem-solving literature (e.g., Newell & Simon, 1972). There, the search for the solution of a problem is described in terms of a search tree whose nodes need to be generated by the problem solver. A tree is a more restricted structure than the graph we have introduced above: it is a directed acyclic graph in which each node has exactly one incoming edge, with the exception of the root node which has no incoming edges. Search proceeds by traversing this search tree. Either a visited node is a goal node, or the next node to be visited needs to be determined. This choice can be made in "systematic" ways that are independent of any domain-specific information. Such blind search techniques rely solely on the positions of the nodes in the search tree. Blind search techniques are designed to search a tree exhaustively. The problem is that, even for a modestly-sized tree, finding a solution may take very long. The choice of the next node to be visited can also be made dependent on domain-specific information like the content of the nodes, and the task at hand. Such heuristic search techniques often reach the goal after a small number of steps, but they are not guaranteed to find a goal node, even if one exists. The best-known blind search techniques are depth-first and breadth-first. In depth-first search, each possible path in the search tree is followed to the end. If Visualizing Individual Differences 7 no solution is found, backtracking to a higher level occurs, and another path is followed to the end. In breadth-first search, each level is searched completely before nodes at the next level are examined. So blind search can be described by simple algorithms. If an evaluation function can be formally specified which ranks the nodes that are candidates for the next visit, heuristic search too can be described by an algorithm. For example, in the popular hill-climbing heuristic search technique, a set of candidates is first evaluated in terms of each node's likely contribution to the goal, before the node with the highest value is chosen. However, as we will discuss in the next section, both the presence and the nature of an evaluation function in human search behavior may be difficult to identify. One of the main goals of a search is to reach a goal quickly, i.e., to minimize the effort caused by the generation and evaluation of nodes. Quick searches usually involve a small number of nodes generated. However, in a rich information space, another goal may be to acquire extensive knowledge about this space. While the Web server logs only give us data about the sequence of pages seen and search terms entered, i.e., navigation paths, we can use these data, with the appropriate caution, to make inferences about the knowledge structures constructed. The extent of navigation preferably combined with response latencies (how many pages were seen, how much time the user spent viewing them) can serve as a rough indicator for how much of the information was "covered" (Schellhas & Brenstein, 1998). Visualizing Individual Differences 8 Web Searches How can general search theory be applied to the description and analysis of searches on the Web? First, the notion of a Web search should be explicated. In general, any traversal of a set of Web pages, in search of specific information, qualifies as a Web search. A more specific meaning is the attempt to find Web pages relating to one or more keywords through the use of a search engine like Google (http://www.google.com) or of a Web index like Yahoo (http://www.yahoo.com). Search engines prompt the user for one or more typedin keywords, and return a set of pages deemed relevant because of syntactic analysis (occurrence of the keyword(s) in the full text or in the HTML meta tags, evaluation of links pointing to this page as in Google’s PageRank method, etc.). Web indices typically offer the user a hierarchical menu of categories (like "education", "sports", "real estate", etc.), leading to lists of hyperlinks deemed relevant to the category by human experts. Web portal is another term used for Web indices, usually with a focus on specific content. The differences between search engines, Web indices, and Web portals are fading: most of today's search engines also offer category menus, and most indices and portals offer at least some keyword specification and full-text search. Users employ both keyword specification and looking through linked lists of pages. All of these entry points into the Web are largely meta servers, linking to content pages distributed across servers all over the world. A central factor of Visualizing Individual Differences 9 their usefulness is that they organize the highly heterogeneous material they point to, and often describe the linked resources in sufficient detail for a survey view of the topic of interest. For example, the German Education Server, used in the study described below, offers a detailed description of each of over 9000 online teaching/learning resources from its database. Thus, it is often easier to find relevant details about a resource on the German Education Server description page than on that resource's own pages. This shows that Web searches may, but do not have to, extend across several servers. A Web search may remain completely within one server, depending on the size and comprehensiveness of that server, and the nature of the information sought. Web searches within one or across several servers can be represented and analyzed in the same manner: in either case, they consist of a sequence of URL requests that may be modeled in the ways described below. This is because in each case, search operates on a Webset. A Webset is a set of pages potentially relevant for the information sought. The Webset may be a list of pages returned by a search engine, plus the search engine's page containing that list, which may be used for refining the search terms, etc. The Webset may also be the home page or a deeper-level page of an index or of a portal, plus all the pages reachable from it. All Websets are subgraphs of the whole Web, and therefore share the same basic structure. The technical caveat is that Web searches within one server can be reconstructed from that server's log files, while Web searches across several Visualizing Individual Differences 10 servers require either the integration of several Web servers' log files (difficult if different organizations host these servers), or require data recording by the client or by a proxy server (see Srivastava, Cooley, Deshpande, & Tan, 2000, for a survey of data collection methods). What happens during a Web search? The general description of the procedure of searching for a solution to a problem can be transferred to analyses of Web navigation. The initial node is the Web page a user starts from. This is typically the home page of the server chosen as entry point, whether this is a search engine, Web index, or portal. The goal node is a node where the user may stop because she feels that she has enough information to answer her initial question. All moves to another Web page are possible moves. They are usually initiated by (a) the activation of a hyperlink or a form submit button, (b) the browser back button, or (c) the browser history. Less frequent are the use of bookmarks, or the typing of a URL (Tauscher & Greenberg, 1997). In their pure forms, blind search techniques are--for any but the smallest Websets-employed mainly by programs, such as Web robots. Most human Web searches are likely to employ a combination of blind and heuristic search techniques. On the one hand, the likely contribution of an option for the goal is assessed: a hyperlink is clicked on because its anchor text appears to be semantically closest to, or at least related to, the topic of interest. A keyword or set of keywords is typed in that appears relevant to the topic of interest. Formalizing this assessment Visualizing Individual Differences 11 of likely contributions amounts to specifying an evaluation function. An evaluation function of Web pages could be similar to the page ranking employed by search engines, where rank increases if the original search terms occur in the Web page’s URL, title tag, etc. So because human Web searches are directed towards a goal, and the descriptions of the next pages reachable from one page are often informative enough to serve as a basis for ranking, some heuristic search will usually be employed. However, sometimes it is difficult to evaluate and rank the different contributions of possible next nodes. In these cases, a user can only choose the next node at random, or resort to a systematic, domain-independent traversal of the available options, i.e. employ blind search techniques. Also, evaluating the merit of a page for one's goals may take longer and/or require more effort than clicking on the hyperlink, waiting for the page to load, reading it, and, if it is not the right one, clicking the back button. This is known as the "base-level/metalevel tradeoff" in AI search theory (Ginsberg, 1993, p. 28). The same idea is expressed in the "optimal information foraging theory" that applies biological modeling of animals' food-foraging to the search for information on the Web (Pirolli & Card, 1995). This tradeoff leads to (at least partial) reliance on blind search techniques. This behavior may be affected by individual differences in cognitive style. For example, Dufresne and Turcotte (1997) have shown that fieldindependent users first browse to gain a survey view of a Webset and then return Visualizing Individual Differences 12 for detailed reading, while field-dependent users seem to follow a more sequential path and read fewer pages. First gaining a survey view involves elements of breadth-first search, while descending down a branch first involves elements of depth-first search. Here, "involving elements of" means that the first strategy will usually not involve every link from the start page, and the second strategy will not descend until the end of one or even all branch(es). In the following, we will refer to such sequences as breadth/depth-first search patterns. (Tauscher & Greenberg, 1997, distinguish depth-first and hub-and-spoke strategies, which correspond to this distinction.) A further reason for concentrating on these patterns is the problem of formalization. Different data collection methods (e.g., video recording, thinkaloud protocols, interviews) and analysis methods (e.g., qualitative content analysis) are needed to find out whether a user has evaluated nodes to be visited next, what the evaluation function has been, and whether she has really picked the node with the highest value. In other words, it will often be difficult to formalize the reasons for regarding one candidate node (one page linked from the current one) as more likely to lead to the goal than another one. Therefore, a formal approach to the analysis of Web navigation should start by concentrating on the structurally identifiable navigation patterns. This kind of analysis can be employed regardless of whether the Webset involves pages from one or from several servers. A pattern in a user’s navigation Visualizing Individual Differences 13 history may involve nodes from one server, or it may contain nodes from several servers. A user may also repeat the same pattern on different servers, for example, to try to gain a survey view first on each new server. Search and the Spatial Metaphor In a search tree as defined above, the main movement is "forward," towards the goal. In physical (and electronic) space, movements are in various directions, often returning to previously visited locations which serve as useful starting points for exploration. These locations are known as landmarks. Knowledge about a space may extend beyond landmarks to encompass the routes one has already traveled, or to a survey representation of the entire space, which allows one to also infer and take shortcuts. The evolution from landmark via route to survey knowledge is observed both in the cognitive development of children and in the development of adults’ knowledge of a new physical space (Siegel & White, 1975). This taxonomy is also applied to Web spaces (e.g., Chen, 1999; Dillon, McKnight, & Richardson, 1993). Landmarks in physical as well as in electronic spaces have three aspects (Sorrows & Hirtle, 1999): visual, cognitive, and structural. Visual refers to the visual uniqueness of the location, cognitive to the meaning, and structural to the importance derived from role or location in the structure of the space. Visualizing Individual Differences 14 The spatial metaphor suggests using visualizations, "maps" of the terrain navigated. We will show next that this also allows one to capture individual behavior in more complex ways than by the measures mentioned above. Patterns in Navigation Behavior A pattern is a regularity in behavior, formulated as a valid expression in a knowledge representation scheme (Witten & Frank, 2000, p. 57). Data are examined with respect to whether they match a pattern template or not. Of particular interest here are sequential patterns. Such patterns can be identified in exploratory analysis (e.g., Olson, Herbsleb, & Rueter, 1994; cf. Spiliopoulou,1999, as an example of the analysis of large-scale Web usage data). They can also be identified in confirmatory analysis (e.g., Fu, 2001). Pattern templates are often described in grammatical form, e.g., by specifying the identities of the first and the last node in a sequence of requests. Pattern knowledge representation schemes can treat the underlying data as sequences or as graphs. In the present study, we decided to formulate a number of grammatical descriptions as minimal versions of the patterns we investigated. Specifically, we investigated: • Breadth-first search patterns: for each node A, we identified sequences [A,B1,A,B2,...,A,Bk], with all visits being first visits and k ranging from 4 Visualizing Individual Differences 15 to 6. I.e., we determined the number of breadth-first search patterns that investigated at least 4 (5,6) nodes reachable from a given node. • Depth-first search patterns: for each node A, we identified sequences [A,B,C,D,...], with all visits being first visits and the number of nodes ranging from 4 to 6. I.e., we determined the number of depth-first search patterns that investigated at least 4 (5,6) new nodes reachable from a given node in an uninterrupted chain. • Landmarks: A node was classified as a landmark if it had more than 5 incoming edges, and these incoming edges were not all from neighboring nodes, i.e. the node was not solely the start of a breadth-first search pattern. Specifically, we required at least one pair of incoming edges to originate from nodes that were at least 5 steps apart in the counts of their first visits. • Routes: for each node A, sequences were identified that were a depth-first search pattern for 2,3, or 4 steps, and then returned by reversing this route. • Survey: This counted the number of moves between nodes that were at least 5, 10, or 15 steps apart in the counts of their first visits. However, these descriptions are only first approximations to the phenomena we are looking for. The use of visualizations allows for more flexibility in the specification of patterns. Using visualizations along with quantitative measures is pervasive in exploratory data analysis (cf. Tukey, 1977). Visualization combines Visualizing Individual Differences 16 two advantages: It allows the representation of large and high-dimensional data sets in a compact space, and it allows the analyst looking at that visualization to survey that space and detect complex, higher-order patterns that often remain unnoticed in a statistical analysis based on numbers (Card, MacKinlay, & Shneiderman, 1999). 6 The attempt to completely formalize expected behavior may even impede exploratory analysis (Olson et al., 1994). Below, we will demonstrate an approach combining measures designed to characterize a navigation history as a whole, with a more localized pattern analysis relying heavily on visualization. An essential question remains: What should be contained in the visualization, and how should it be visualized? Visualization of Navigation Paths Visualization of paths usually occurs either within graphs depicting the whole site, in tree shape, or as a diagram plotting URLs against time. In the visualization of whole sites, the size of the graph is the central challenge. The most common criteria for a well-drawn graph are minimized edge crossings, uniform edge lengths, straight edges and symmetry. Depending on the application, nodes should be evenly distributed or distributed in space such that nodes with stronger connections are closer to one another, and nodes with weaker connections are farther apart (Chen, 1999). To reach these goals, the number of edges may need to be reduced, clustering algorithms be applied to group semantically linked nodes, nodes and edges may have to be displayed in a deliberately distorted way (e.g., Visualizing Individual Differences 17 explicit or implicit focus-and-context views), and interactive techniques may have to be applied to help the user compose a mental representation integrating several perspectives (e.g., Munzner, 1998; for a recent survey, see Chen, 1999). This basic approach can be applied to the display of navigation graphs. For example, Chen (1999) computes the transition probabilities between any pair of nodes and uses these as connection strengths. The resulting data can be displayed analogously to a Web site linkage structure. This can show the perceived proximity or semantic closeness of two nodes well, assuming that they are often visited in sequence. However, a transition probability summarizes all transitions between two nodes, ignoring the history of navigation around this transition. This makes them unsuitable for an in-depth analysis of navigation sequences (see also Niegemann, 2000). Graphs that depict movements in the context of the whole site help to detect, for example, available choices that are never taken, but make it difficult to interpret navigation patterns by content, because the placement of nodes in space is governed by the need to distribute all the site's URLs across this space. Spiliopoulou (1999) uses trees generated from a set of different users’ navigation paths that share, e.g., the same start and end nodes. This means that every single navigation path is in effect displayed as a left-to-right chain of nodes and edges. While this gives a good idea of the behavior of large groups of users, it relies on textual annotations (each node is annotated by the Web page name and the Visualizing Individual Differences 18 occurrence number within the navigation paths considered). Thus, important aspects of navigation structure such as the recurrent visits to nodes, are not easily seen. Also, only the position of the nodes along the x axis is meaningful (steps within the navigation paths). The position along the y axis cannot be interpreted. Hemmje, Kunkel, and Willett (1994) derive semantic hierarchies of Web pages from the graph representation of the Web site and display the actual navigation history against this "skeleton" of content-oriented search paths. The hierarchies are trees, and they are visualized as three-dimensional cone trees, which provide an implicit focus-and-context view. While this is a good support for users (helping them orient themselves and navigate by supplying a semantics-based "you-are-here map"), it is less helpful for the analysis of navigation paths. Repeated visits to one node do not stand out in the visualization, and a possible structure generated by the user’s interaction with the hypertext cannot become visible, because navigation is framed within the structure of the Web site as perceived by the content designer. Another approach is plotting of URLs against time (e.g., Hasebrook & Fezzardi, 1996; Schellhas & Brenstein, 1998). However, time plots can lead to erroneous interpretations of the shape of the graph, which is often meaningless because URLs are arranged in an arbitrary order along the y axis. In some cases, URLs can be ordered along the y axis in a meaningful way (Berendt, 2000; Niegemann, 2000). However, this is not possible for all types of sites or parts of Visualizing Individual Differences 19 sites, and patterns involving revisits of nodes are difficult to see. These techniques are therefore less appropriate for supporting exploratory analysis of navigation in parts of a site. Requirements for the Visualization of Individual Navigation Graphs and their Realization in STRATDYN The above discussion has shown that existing approaches to the visualization of Web navigation pose problems for personand task-specific analyses. In the following, we have therefore formulated requirements for visualizing individual navigation graphs which will be described with reference to the tool STRATDYN. This new approach to the analysis and visualization of log file data enables researchers to compute a number of quantitative measures of navigation behavior, while at the same time offering a concise visual representation of navigation episodes within a single session. The visualization is generated in two stages. In the graph definition stage, the ordered list of URL requests is transformed into a navigation graph as defined above. Moreover, several quantitative measures (such as number of nodes and edges, the number of backward moves and back button moves, and branching at home page) are generated. In the graph layout stage, the graph is drawn, with configuration parameters determining how many and whose search episodes are aggregated into one graph, whether branching at the start page or branching at lower-level pages is emphasized, etc. The resulting navigation graphs can display Visualizing Individual Differences 20 navigation histories of one or several users. In the analyses reported below, we introduce individualized site maps that show all navigation moves of one user, emphasizing branching at the start page. Visualization requirements are translated into the use of the visual variables (Bertin, 1983; Mackinlay, 1986) shape, position, and color, each encoding a dimension of the data. This small number of visual variables generates additional structure in the visualization by encouraging perceptual grouping. It also allows patterns to emerge, as will be discussed in the results section. Number of nodes and extent of navigation A first requirement for visualization concerns the number of nodes. The number of nodes in most experimental navigation tasks, like ours, is small compared to even a modestly sized Web site's number of nodes. This means that scalability to very large graphs and algorithmic efficiency are of lesser concern in psychological research. In STRATDYN, the visual variable shape is used to encode whether a graphical element is a node or an edge, and the identity of a node. Each node is visualized by its name. Each movement between two nodes is shown as an arrow. The resulting image can be treated as a survey view, which enables researchers to readily obtain a global view (targeted vs. disoriented navigation) (see Figure 1 for an example). The relative size of different navigation graphs encodes extent of navigation. At the same time, in an enlarged view, paths can be traced and utilized as a basis for content analysis (see Figure 3 for an Visualizing Individual Differences 21 example). This tracing of paths, which requires reading the visited pages’ names, is not necessary to recognize and understand the extent and structure of the search. For this stage of analysis, individualized site maps can therefore be scaled as shown in Figures 1, 4, and 5. Insert Figure 1 about here Meaningful dimensions A second requirement are meaningful dimensions in xy space. By this we mean that the dimensions of the graph drawing, i.e. the relative positions of elements in the drawing's xy space, should be interpretable. This is necessary because humans tend to interpret the ordering of magnitudes in a diagram as corresponding to an ordering in reality (e.g., Berendt, Rauh, & Barkowsky, 1998). In STRATDYN, the x dimension (left-right) encodes temporal order within the processing of one task, or the temporal order of visits to one node. Figure 2 shows a fictitious example of 7 consecutive moves. The arrows should be read as follows: The first outward move from a node i to a node j is shown by a straight left-to-right arrow if this movement is the first inward move into j (movements (1) and (5)). Otherwise, the movement is shown by the leftmost outgoing edge of the node i (movements (2) and (6)). An outgoing edge has no arrowhead at i. The next movement is either another straight left-to-right arrow, or the edge Visualizing Individual Differences 22 immediately to the right of the previous move’s incoming end (the end with an arrowhead) (movements (3), (4), and (7)). To improve visual clarity, backward moves are also distinguished from forward moves by line width. Insert Figure 2 about here The y dimension encodes temporal order as well. First, it (redundantly) codes temporal order of multiple moves to and from a node (cf. Figure 2, where segments of the edge showing movement 4 are to the right of and below movement 3, and similarly for movement 7 in relation to movement 6). Second, the y dimension encodes the temporal order of the experimental tasks. Representation of moves: exhaustiveness and selectivity A third requirement is that the graph drawing should combine the advantages of trees with those of full graphs, i.e., visual clarity and information richness. In the problem-solving-as-search literature, the traversed graph is a tree. A given node can be the child node of several other nodes, and if it is generated several times in this way, it is duplicated. A graph drawing that duplicates nodes, however, obscures the interconnectedness of problem solving, the re-use of already explored nodes and subgraphs. The selective re-entry into and exit from previously seen sequences is crucial for understanding the use of a non-linear text. Nodes should not be duplicated. Another common class of tree representations, Visualizing Individual Differences 23 spanning trees, do not duplicate nodes but eliminate edges to enhance visual clarity (Munzner, 1998; for pathfinder networks as an alternative way of eliminating edges, see Chen, 1999). A spanning tree of a graph maintains, for each node, exactly one path from a previously chosen root node to that node. This elimination of edges increases visual clarity, but it takes away information. Munzner (1998) lets users control the number of displayed edges interactively, thus making that information available. In STRATDYN, a spanning tree is used as the visualization’s basic layout, but at the same time, the full graph is displayed. The visualization offers three independent ways of perceptual grouping. The first grouping shows the nodes in the order of their first visits: the criterion for positioning a node, which is also the criterion for the generation of the spanning tree, is the first visit to a node. This criterion emphasizes the discovery of information. The grouping by task is created by the arrangement in parallel horizontal rows and by similarity of shape: the nodes that are arranged in this basic layout are strings of letters. The second grouping shows all movements, which includes single and recurrent movements that are independent of the order of first visits to the nodes involved. This emphasizes navigation. This grouping is created by similarity of shape: the edges are dominated by horizontal and vertical straight lines (the arrowheads are comparatively small and restricted to the end points of edges, so they do not affect the grouping of edges noticeably). Visualizing Individual Differences 24 Representation of nodes and edges: task structure The third way of grouping by the visual variable color encodes the experimental task (task 1: red, task 2: green, and so on). In halftone printing, the colors appear as different gray values. This is needed as a way of grouping nodes and edges that is independent of the grouping by temporal order: first, each new task involves the opening of a new horizontal branch, but not every new horizontal branch indicates the start of a new task. Second, repeat visits to nodes are particularly interesting if they occur in new tasks, since this indicates possible transfer. Patterns A final requirement is the visibility of patterns. The individual moves must be visible, but in addition, we want to be able to identify patterns, and to compare their realizations in different navigation graphs. The patterns we are looking for are: breadth-first search patterns, depth-first search patterns, landmarks, routeoriented navigation, and survey-oriented navigation. After the identification of patterns, the analyst can proceed to investigate these behavioral sequences qualitatively. Application of the STRATDYN Approach in Research on Individual Differences in Navigation Behavior How this approach works in practice is best illustrated by example. In the following, we therefore show how STRATDYN can be used to visualize different Visualizing Individual Differences 25 patterns of navigation in a multi-task condition, and to provide meaningful measures for the quantitative analysis of log file data. To this end, we present selected findings from two studies in which we addressed the following questions: Do individual differences in the ability to focus attention on relevant visual stimuli affect (a) the quality of the interaction as expressed in different navigational patterns, (b) the extent of navigation, i.e. the number of pages visited, and (c) the effectiveness of navigation (number of correct responses in study 1, or the amount of relevant information gathered in study 2). Following a brief description of the method, we present selected findings to illustrate how the visualizations and measures provided by the STRATDYN approach can aid in answering these questions. A more thorough description of the studies can be found in (Brenstein & Berendt, 2001).