Model-based Navigation Support Herre van Oostendorp Department of Information and Computer Sciences Utrecht University herre@cs.uu.nl In collaboration with Ion Juvina Dept. of Psychology, Carnegie-Mellon University Presented at Seminar Computational Cognitive Modeling of Web Navigation 19-20 March, 2010 Utrecht University 1
Overview 1. Cognitive modeling of web navigation CoLiDeS+ 2. Studies (2x) on model-based navigation support 3. Conclusions & Discussion 4. Practical relevance 5. Limitations 2
1. Cognitive modeling of web navigation A model describing web navigation is CoLiDeS+ (Juvina, 2006; Van Oostendorp & Juvina, 2007), extension of: CoLiDeS, Comprehension-based Linked model of Deliberate Search (Kitajima, Blackmon & Polson, 2000) Successfully used in repairing usability problems Information seeking and navigation is driven by text comprehension and problem solving processes, guided by users goals and information scent (semantic similarity between object and goal of the user (Pirolli & Card, 1999)) Acting on a page screen object, the outcome of multi-step process parsing current display focusing on screen objects comprehending screen objects selecting screen object with highest semantic similarity to user s goal We focus on the last step 3
Information scent as driver of navigation behavior Information foraging theory (Card, Pirolli et al., 2001) (metaphore: animals finding food locations in a wood on basis of promising trails or scent ) On basis of information scent users try to find information patches ; scent is driving behavior. Information patch concentration of information relevant to goal. Information scent - perceived compatibility/similarity to a goal. Often measured with Latent Semantic Analysis (Landauer et al., 1998): semantic similarity 4
CoLiDeS: central idea scent (LSA) Web page title Heading 1 Heading 2 Link description 1 Link description 3 Link description 2 Link description 4 LSA=? Goal descriptio n Semantic similarity is computed with Latent Semantic Analysis technique (LSA). LSA technique is similar to factor analysis; the meaning of a word, sentence or text is represented as a vector in a high dimensional space. The degree of similarity is measured by the cosine value (correlation) between corresponding vectors. 5
CoLiDeS+: path adequacy includes contextual information by computing path adequacy, based on LSA (next to scent ). Path adequacy is the semantic similarity between the succession of links that have been selected to a particular moment in a navigation session and the goal (computed with LSA) past page past page past page Link label 3 Link label 2 Link label 1 Goal LSA=? current page Link label 4 6
Characteristics of CoLiDeS+ Navigation behavior of CoLiDeS+ based on Information scent (semantic similarity link label and goal) Path adequacy (semantic similarity navigation path and goal) Navigation strategies as next best strategy, refocusing and backtracking Details Van Oostendorp & Juvina, 2007 7
Diagram of the algorithm that implements the CoLiDeS+ model User s goal Scent (LSA) Path adequacy Strategies 8
Example CoLiDeS+ algorithm I want to find a hotel as soon as possible and as cheap as possible Sleep LSA.32 * Eat LSA.27 Sign up for a waiting list.73 * Book accommodation now.44 Registered customer.10 New customer.25 PA Sleep, sign up for waiting list PA:.72 Sleep, sign up for waiting list, new customer PA:.61 path switch Select next best and click on Book accommodation now * Economical rooms.36* Exclusive apartments.24 Sleep, book accommodation now, economical rooms PA:.55 Sleep, book accommodation now PA:.51 (though lower (.36) than.44, PA is increasing from.51 to.55). Thus, click: economical rooms * 9
2. Studies on Model-based Navigation Support Based on the model CoLiDeS+, and applying the algorithm, it is possible to determine at each step in a simulation of a task what is the successful path. At the end we have together the successful path consisting of relevant links leading to the target information. We can use this information to generate navigation support and to reduce navigation problems Hypothesis: the model-generated support improves navigation and performance, particularly for users with low spatial ability (There is a high correlation between spatial ability and path adequacy, Juvina 2006). 10
Study 1 Voice suggestions Support by voice suggestions Link suggestions: Click on <link label> Path switch suggestions: Go back Method 29 subjects (students) 6 realistic web tasks and associated web sites (Morrison et al., 2000) E.g. you want to buy a new car but you don t have enough money, calculate how much you can loan, and what your monthly payment will be (given certain assisting web sites ) Tasks demanding a combination of searching information and problem solving/decision making Instruction is the basis of the goal description 11
Support The tasks were simulated with Colides+ prior to the actual navigation session The results of simulations were successful paths, i.e. successions of links leading to target pages. Based on these results, above link suggestions such as Click on were given. It was told that the suggestions were meant to help, not mandatory to follow. Task performance (mean correctness over 6 tasks on a scale from 0-4) Spatial ability (mental rotation task, Shepard & Metzler, 1971) 12
Results Mean task performance Task performance 1,4 1,2 1 Spatial Ability Increased task performance with Support Compensation for deficit of spatial ability 0,8 0,6 0,4 0,2 0 Control Support Low High Navigation: number of steps sign. less in the support condition, and less time needed. 13
Results Furthermore CoLiDeS+ simulated users behavior better than CoLiDeS We compared what selections users made with the selections made by CoLiDeS and CoLiDeS+ respectively User selections: for each user, for each task, for each visited webpage we recorded which options users selected. We ran CoLiDeS and CoLiDeS+, and calculated the selections of each model User selections can match or not match the model selections. For (10 control) subjects we found Total valid selections: 275 CoLiDeS selection matches user selection: 129 46.9% CoLiDeS+ selection matches user selection: 151 54.9% (Х 2 =3.52, p<.06) Still subjects did not like the auditory modality (annoying/disturbing) 14
Study 2 Graphical suggestions Method Presentation of graphical suggestions, based on the successful paths during the simulations, in the form of red arrows 15
Hypotheses: Model-generated support improves User s perception (feeling of disorientation) Navigation (more direct) Task performance (correctness) 16
Method Subjects (students): control condition n=16, support condition n=14 Materials: same 6 realistic web tasks and associated web sites E.g. you want to buy a new car but you don t have enough money, calculate how much you can loan, and what your monthly payment will be (given certain assisting web sites ) Tasks demanding a combination of searching information and problem solving/decision making Instruction is the basis of the goal description Dependent variables perception of lostness : disorientation scale of Ahuja & Webster, 2001 navigation behavior task performance (mean over the 6 tasks, score range 0-4 per task) 17
Results Sign. less disorientation, particularly for men, in support condition Mean Perceived Disorientation 1,8 1,6 1,4 1,2 1 women 0,8 man 0,6 0,4 0,2 0 Control Support Positive perception of suggestions Sign. more linear navigation in support condition 18
Results Mean task performance 3.5 3 2.5 2 women men 1.5 1 control support Sign. increase in task performance in the support condition Effect of gender was not significant 19
3. Conclusion & Discussion Model-generated navigation support is useful with graphical suggestions also positive effects with auditory support Inherent browsing tasks, not simple information retrieval tasks (such as search engines); preserves user experience (relevant contextual information) Discussion Specific advantages of model-generated support: Goal specific: it can be made adaptive, i.e. suggestions tailored to the goals It opens the possibility of automatization Useful in the case of cognitive or perceptual limitations; Dynamic (running few steps ahead; not fully implemented) No need to improve design of documents 20
4. Practical relevance Case: navigation support for vision impaired users Screen readers Cognitive overload Selective reading, e.g with model-generated suggestions Pilot study simulated VIP s screen off- screen reader- mechanism with suggestions based on reading priority (relevant links first). no effects of task performance or navigation, though suggestion mechanism was effective (in support condition relevant links were successfully pushed up). Maybe break in coherence or manipulation not strong enough More general: situations that require handling information overload with limited perceptive and cognitive capabilities (e.g. multitasking persons). This model-based support is probably in the end more useful and worthwhile than well-known heuristics (Nielsen), and guidelines, because it is more precise and theory-driven. 21
5. Limitations However, at this moment there has still work to be done, e.g. how can pictorial information be included in the model (see the contribution of Saraschandra) It still needs automatization! 22
Thank you herre@cs.uu.nl 23
Main References Ahuja & Webster (2001). Perceived Disorientation; an Examination of a New Measure to Access Web Design Effectiveness. Interacting with Computers, 14(1),15-29 Card, Pirolli, Van der Wege, Morrison, Reeder, Schraedly & Boshart (2001). Information Scent as Driver of Web Behavior Graphs: Results of a Protocol Analysis Method for Web Usability. In SIGCHI 2001 Proceedings. Seattle: ACM Press. Juvina (2006). Development of a Cognitive Model for Navigating on the Web. Dissertation Utrecht University. Kitajima, Blackmon & Polson (2000). A Comprehension-based Model of Web Navigation and Its Application to Web Usability Analysis. In People & Computers XIV, Springer. Landauer, Foltz & Laham (1998). Introduction to Latent Semantic Analysis. Discourse Processes, 25, 259-284. Morrison, Pirolli & Card (2000). A Taxonomic Analysis of What World Wide Web Activities Significantly Impact People s Decisions and Actions. Xerox PARC. Pirolli & Card (1999). Information Foraging. Psychological Review, 106, 643-675. Van Oostendorp & Juvina (2007). Using a cognitive model to generate web navigation support. International Journal of Human-Computer Studies, 65, 887-897. Shepard & Metzler (1971). Mental Rotation of Three-Dimensional Objects. Science, 171, 701-703 24