HCI Forum Topic DISS 720
from www.pixelcharmer.com
Cognitive Models for Web Design
Information Foraging Theory Applied
May 2002
Information foraging theory seeks to explain information-seeking behavior in humans. Central to its thesis is that information foraging is an exaptation of food foraging mechanisms, therefore models of optimal foraging theory developed by anthropologists and ecologists in the study of food foraging will help us understand foraging behavior in consumers of information. These models allow us to investigate foraging behavior in relation to particular environmental conditions and the constraints of foraging for information in a dynamic ecology.
Information foraging theory gives those researching user interaction with Web sites a way to examine user goals, their decision making processes and adaptations to the information access system environment. Researchers can then make use of this knowledge in assessing system and interface design. Most importantly to those charged with developing a web site, information foraging theory can then inform design. I will demonstrate and give examples of ways web developers can use information foraging theory to cultivate more attractive paths to richer patches of information on a web site by knowing their visitors' information diets, allowing users to take advantage of the paths created by others, and providing representations of content with a strong information scent.
New Set of Tools
Users assess the appropriateness of following a particular path on the Web by considering a representation, usually a textual description or graphic, of the distal content. Furnas (1997) explained that a representational object held a “residue” of what lay behind it. Residue was recast and refined by Pirolli (1997) as information “scent” and defined in Card et al. (2001) as a user’s “(imperfect) perception of the value, cost, or access path of information sources obtained from proximal cues, such as WWW links.” In the initial work by Pirolli and Card (1995) on information foraging, they defined the profitability of an information source “as the value of information gained per unit cost of processing the source.” Cost is defined in terms of time spent, resources utilized and opportunities that are lost when pursuing another particular strategy instead of others. (Russell, 1993)
In order to invent a new set of tools for informing design of information systems, Pirolli and his colleagues went on to develop a computational cognitive model of information foraging based on ACT-R. The originator of ACT-R, John Anderson, used a network model of knowledge to develop his architecture. It solves the network model problem of defining associations among nodes by representing knowledge in a proposition. Therefore, the ideas in a proposition reveal their relationships to each other by their placement following linguistic rules within the proposition. When one node is activated in the network model, then a related node is activated as well and so on, spreading activation among related nodes. As with other network models, where to stop with this spreading, or “degree of fan” is problematic. However, ACT-R is very useful for modeling user interaction in a task environment. (Reisberg, 2001 pp.253-262)
Pirolli also discussed an overall framework for studying human-computer interaction from an ecological and cognitive perspective by reiterating the levels of analysis for understanding an information processing system. Unlike Marr (1982) he breaks up the first level “what the device does and why,” so that his levels number four in total. The first level is adaptation, then knowledge, followed by the cognitive level and finally biological or the implementation level as termed by Marr. (Pirolli, 1997)
Within this structure, Pirolli developed the “adaptive control of thought in information foraging (ACT-IF)” to model optimal foraging in a large collection of texts. In particular, using the Scatter/Gather browser interface developed at Xerox PARC, they were able to model users following information scent. Spreading activation could be measured starting from a task query to the relevant information. The Scatter/Gather browser would communicate the contents of the text collection by clustering the topics of each into discrete related groups represented by snippets of text. Following the rules set forth in ACT-IF, one or more clusters were selected to be scattered (reclustered) in the Scatter/Gather browser into topically related groups until the task was complete. When ACT-IF could make accurate judgments about distal information, thereby activating the nodes from information goal to that piece of distal information that completed the task, the proximal representation was considered to have strong information scent. (Pirolli, 1997)
ACT-IF allows (simulated) users with different constraints to be tested interacting with variations on a design. Using ACT-IF can allow a greater number of design variations to be tested under more conditions than in traditional user testing. Comparing the results from actual user tests with similar tasks performed by ACT-IF can test its accuracy. In fact some comparisons were made, but they are few due to the laborious and time-consuming nature of handcoding each of the results from videotaped user tests. (Pirolli, 1997)
Assuming that real users will strive for the optimal foraging behavior seems at odds with the frequently observed problem solving strategy know as “satisficing.” In fact, the process of making decisions based on aspiration level seems to provide a better description of activity observed in real world user testing. (Krug, 2000, p.24) However, Pirolli briefly points out that “satisficing can often be characterized as localized optimization (e.g., hill climbing) with resource bounds and imperfect information as included constraints.” (Pirolli, 1999 p.645) In addition, David Ward, et al. examined the role of satisficing in food foraging theory and found it wasn't at odds with optimal foraging theory. (Ward, 1992)
ACT-IF is a very useful tool for examining possible designs for a large web site composed of many individual texts. However, it’s efficacy with collections of images and non-text representations of distal information has not been considered.
Another tool that has implications for design allows us to analyze user paths from information in web server logs. Although there are many pieces of software for computing statistics from web server logs, none of them allows us to extrapolate user goals. Pirolli and his collegues demonstrate a way to take surfing patterns and infer the associated information need of a given user. Users are then clustered together when similar needs are identified. Developers can then construct user types, or “user profiles” for a particular site. (Chi, 2001, Heer, 2000)
Inferring User Need by Information Scent (IUNIS) was the algorithm that allowed the development of a tool for building user profiles from surfing patterns. IUNIS identifies the documents that a user accessed during a browsing session and the order they were accessed. Applying the longest repeating subsequence (LRS) assists in extracting paths that are repeated by multiple users, and therefore more likely to be relevant to our task. Each of these repeated paths help us to describe a user profile. Vector distances between pages are calculated and distances between vectors are as well. Four modalities are then identified for each web page accessed in the path so we may cluster them.
1. each unique word in a page (however it is weighted as less significant if the word is found frequently in other pages on the site)
2. the directory location of that page as represented by forward slashes in its URL (page is given more weight if fewer other documents share the directory)
3. how many links from other pages on our site point to that page (weighted so that a link from a particular page is less significant if the same page points to several others)
4. all the links that go out from our page whether they only link to other pages on our site or not.
Once each of these modalities and vectors are identified for all of the pages within statistically significant paths, we have what is known as the CUT data of a site. Before completing our calculation, we weight the final pages in a path, or in other words, the pages more recently accessed so as not to give too much importance to gateway pages or splash pages that everyone must click through. We now have a representation of our site by multi-modal vector paths. We can cluster our pages and unlike prevalent web log analysis software that only analyzes one mode (number of hits to a page, number of links to a page, etc.) our multimodal representations make it possible for us to construct user profiles from our calculations. (Chi, 2001)
Web site developers design sites for specific user groups as identified by extensive marketing research and if feasible ethnographic study including contextual inquiry. Following the practice of these user-centered design methods, developers then construct hypothetical, archetypal users to build user profiles and inform design. However, by generating user profiles from analyzing the Web server logs of an existing site, future iterations of the site can better meet its users' needs without requiring its developers to conduct actual marketing surveys and contextual inquiries.
A third tool that can be investigated in regard to information foraging theory is collaborative filtering. Collaborative filtering allows users to forage for information in groups much like a group of humans banding together to hunt for food when objects included in their diet are distributed widely and thinly in their environment. By ascribing a history of use to a digital object, a single user can benefit from the foraging of others. Interaction history of other foragers or as described by Wexelblat (1999), “footprints ...allow users to leave traces in the virtual environment...” The interaction history of others, attached to an object can come from passive sources, such as access logs, or active sources, such as online papers that allow users to leave commentary. There are several shopping sites currently using both of these methods. For example, Amazon.com extracts user information from its logs so a single item’s description can also include items that other users viewed or bought when viewing or buying the current item. Amazon.com also uses active interaction history by soliciting user opinions of each product then attaching that interaction to the appropriate item. Potential consumers can then view and benefit from another user’s experience when making their purchasing decision.
As Card and Pirolli hoped, information foraging theory has already provided Web developers with new tools. Among these tools are ways to spread activation by using labels with strong information scent so that paths are more attractive and lead to richer patches of information. By efficiently constructing user profiles developers can know their users' information diet and increase the profitability of items in their diets by decreasing the amount of energy expended when foraging for desirable items. Allowing users to take advantage of the paths created by others through collaborative filtering, as Wexelblat (1999) demonstrated, leads to greater user satisfaction and greatly reduces the cost associated with foraging.
Card, Stuart K., Peter Pirolli, Mija Van Der Wege, Julie B. Morrison, Robert W. Reeder, Pamela K. Schraedley, Jenea Boshart (2001). Information scent as a driver of Web behavior graphs. Proceedings of the Conference on Human factors in computing systems CHI '01 Association for Computing Machinery.
Chi, Ed H. Peter Pirolli, Kim Chen, James Pitkow (2001). Using Information Scent to Model User Information Needs and Actions on the Web. In Proc. of ACM CHI 2001 Conference on Human Factors in Computing Systems, pp. 490-497. ACM Press, April 2001. Seattle, WA.
Furnas, G. W, (1997). Effective view navigation. In Proceedings of the Human Factors in Computing Systems, CHI '97 (pp. 367-374). Atlanta, GA: Association for Computing Machinery.
Heer, Jeffrey Ed H. Chi (2000) Identification of Web User Traffic Composition using Multi-Modal Clustering and Information Scent. in Proc. of the Workshop on Web Mining, SIAM Conference on Data Mining, April 2001, Chicago, IL. pp. 51-58
Krug, Steve (2000). Don't make me think: a common sense approach to web usability. Macmillan USA.
Marr, D. (1982) Vision. San Francisco: W.H. Freedman.
Pirolli, Peter, Stuart Card (1995). Information Foraging in Information Access Environments. In Proceedings of the Human Factors in Computing Systems, CHI '95. Association for Computing Machinery.
Pirolli, P. (1997). Computational models of information scent-following in a very large browsable text collection. In Proceedings of the Conference on Human Factors in Computing Systems, CHI '97 (pp. 3-10). Atlanta, GA: Association for Computing Machinery.
Pirolli, Peter, Stuart Card (1999). Information Foraging. Psychology Review Vol. 106, No. 4. (pp.643-675)
Reisberg, Daniel (2001). Cognition: exploring the science of the mind. 2nd ed. W.W. Norton & Company, Inc.
Russell, Daniel M., Mark J. Stefik, Peter Pirolli, Stuart K. Card (1993). The cost structure of sensemaking. Proceedings of the Conference on Human Factors in Computing Systems. Association for Computing Machinery.
Ward, David, Jacob Blaustein (1992) The role of satisficing in foraging theory. Oikos 63:2 (pp. 312-317).
Wexelblat, Alan, Pattie Maes (1999) In Proceedings of the Human Factors in Computing Systems, CHI '99. Association for Computing Machinery.
Cognitive Models for Web Design
Information Foraging Theory Applied
May 2002
Information foraging theory seeks to explain information-seeking behavior in humans. Central to its thesis is that information foraging is an exaptation of food foraging mechanisms, therefore models of optimal foraging theory developed by anthropologists and ecologists in the study of food foraging will help us understand foraging behavior in consumers of information. These models allow us to investigate foraging behavior in relation to particular environmental conditions and the constraints of foraging for information in a dynamic ecology.
Information foraging theory gives those researching user interaction with Web sites a way to examine user goals, their decision making processes and adaptations to the information access system environment. Researchers can then make use of this knowledge in assessing system and interface design. Most importantly to those charged with developing a web site, information foraging theory can then inform design. I will demonstrate and give examples of ways web developers can use information foraging theory to cultivate more attractive paths to richer patches of information on a web site by knowing their visitors' information diets, allowing users to take advantage of the paths created by others, and providing representations of content with a strong information scent.
New Set of Tools
Users assess the appropriateness of following a particular path on the Web by considering a representation, usually a textual description or graphic, of the distal content. Furnas (1997) explained that a representational object held a “residue” of what lay behind it. Residue was recast and refined by Pirolli (1997) as information “scent” and defined in Card et al. (2001) as a user’s “(imperfect) perception of the value, cost, or access path of information sources obtained from proximal cues, such as WWW links.” In the initial work by Pirolli and Card (1995) on information foraging, they defined the profitability of an information source “as the value of information gained per unit cost of processing the source.” Cost is defined in terms of time spent, resources utilized and opportunities that are lost when pursuing another particular strategy instead of others. (Russell, 1993)
In order to invent a new set of tools for informing design of information systems, Pirolli and his colleagues went on to develop a computational cognitive model of information foraging based on ACT-R. The originator of ACT-R, John Anderson, used a network model of knowledge to develop his architecture. It solves the network model problem of defining associations among nodes by representing knowledge in a proposition. Therefore, the ideas in a proposition reveal their relationships to each other by their placement following linguistic rules within the proposition. When one node is activated in the network model, then a related node is activated as well and so on, spreading activation among related nodes. As with other network models, where to stop with this spreading, or “degree of fan” is problematic. However, ACT-R is very useful for modeling user interaction in a task environment. (Reisberg, 2001 pp.253-262)
Pirolli also discussed an overall framework for studying human-computer interaction from an ecological and cognitive perspective by reiterating the levels of analysis for understanding an information processing system. Unlike Marr (1982) he breaks up the first level “what the device does and why,” so that his levels number four in total. The first level is adaptation, then knowledge, followed by the cognitive level and finally biological or the implementation level as termed by Marr. (Pirolli, 1997)
Within this structure, Pirolli developed the “adaptive control of thought in information foraging (ACT-IF)” to model optimal foraging in a large collection of texts. In particular, using the Scatter/Gather browser interface developed at Xerox PARC, they were able to model users following information scent. Spreading activation could be measured starting from a task query to the relevant information. The Scatter/Gather browser would communicate the contents of the text collection by clustering the topics of each into discrete related groups represented by snippets of text. Following the rules set forth in ACT-IF, one or more clusters were selected to be scattered (reclustered) in the Scatter/Gather browser into topically related groups until the task was complete. When ACT-IF could make accurate judgments about distal information, thereby activating the nodes from information goal to that piece of distal information that completed the task, the proximal representation was considered to have strong information scent. (Pirolli, 1997)
ACT-IF allows (simulated) users with different constraints to be tested interacting with variations on a design. Using ACT-IF can allow a greater number of design variations to be tested under more conditions than in traditional user testing. Comparing the results from actual user tests with similar tasks performed by ACT-IF can test its accuracy. In fact some comparisons were made, but they are few due to the laborious and time-consuming nature of handcoding each of the results from videotaped user tests. (Pirolli, 1997)
Assuming that real users will strive for the optimal foraging behavior seems at odds with the frequently observed problem solving strategy know as “satisficing.” In fact, the process of making decisions based on aspiration level seems to provide a better description of activity observed in real world user testing. (Krug, 2000, p.24) However, Pirolli briefly points out that “satisficing can often be characterized as localized optimization (e.g., hill climbing) with resource bounds and imperfect information as included constraints.” (Pirolli, 1999 p.645) In addition, David Ward, et al. examined the role of satisficing in food foraging theory and found it wasn't at odds with optimal foraging theory. (Ward, 1992)
ACT-IF is a very useful tool for examining possible designs for a large web site composed of many individual texts. However, it’s efficacy with collections of images and non-text representations of distal information has not been considered.
Another tool that has implications for design allows us to analyze user paths from information in web server logs. Although there are many pieces of software for computing statistics from web server logs, none of them allows us to extrapolate user goals. Pirolli and his collegues demonstrate a way to take surfing patterns and infer the associated information need of a given user. Users are then clustered together when similar needs are identified. Developers can then construct user types, or “user profiles” for a particular site. (Chi, 2001, Heer, 2000)
Inferring User Need by Information Scent (IUNIS) was the algorithm that allowed the development of a tool for building user profiles from surfing patterns. IUNIS identifies the documents that a user accessed during a browsing session and the order they were accessed. Applying the longest repeating subsequence (LRS) assists in extracting paths that are repeated by multiple users, and therefore more likely to be relevant to our task. Each of these repeated paths help us to describe a user profile. Vector distances between pages are calculated and distances between vectors are as well. Four modalities are then identified for each web page accessed in the path so we may cluster them.
1. each unique word in a page (however it is weighted as less significant if the word is found frequently in other pages on the site)
2. the directory location of that page as represented by forward slashes in its URL (page is given more weight if fewer other documents share the directory)
3. how many links from other pages on our site point to that page (weighted so that a link from a particular page is less significant if the same page points to several others)
4. all the links that go out from our page whether they only link to other pages on our site or not.
Once each of these modalities and vectors are identified for all of the pages within statistically significant paths, we have what is known as the CUT data of a site. Before completing our calculation, we weight the final pages in a path, or in other words, the pages more recently accessed so as not to give too much importance to gateway pages or splash pages that everyone must click through. We now have a representation of our site by multi-modal vector paths. We can cluster our pages and unlike prevalent web log analysis software that only analyzes one mode (number of hits to a page, number of links to a page, etc.) our multimodal representations make it possible for us to construct user profiles from our calculations. (Chi, 2001)
Web site developers design sites for specific user groups as identified by extensive marketing research and if feasible ethnographic study including contextual inquiry. Following the practice of these user-centered design methods, developers then construct hypothetical, archetypal users to build user profiles and inform design. However, by generating user profiles from analyzing the Web server logs of an existing site, future iterations of the site can better meet its users' needs without requiring its developers to conduct actual marketing surveys and contextual inquiries.
A third tool that can be investigated in regard to information foraging theory is collaborative filtering. Collaborative filtering allows users to forage for information in groups much like a group of humans banding together to hunt for food when objects included in their diet are distributed widely and thinly in their environment. By ascribing a history of use to a digital object, a single user can benefit from the foraging of others. Interaction history of other foragers or as described by Wexelblat (1999), “footprints ...allow users to leave traces in the virtual environment...” The interaction history of others, attached to an object can come from passive sources, such as access logs, or active sources, such as online papers that allow users to leave commentary. There are several shopping sites currently using both of these methods. For example, Amazon.com extracts user information from its logs so a single item’s description can also include items that other users viewed or bought when viewing or buying the current item. Amazon.com also uses active interaction history by soliciting user opinions of each product then attaching that interaction to the appropriate item. Potential consumers can then view and benefit from another user’s experience when making their purchasing decision.
As Card and Pirolli hoped, information foraging theory has already provided Web developers with new tools. Among these tools are ways to spread activation by using labels with strong information scent so that paths are more attractive and lead to richer patches of information. By efficiently constructing user profiles developers can know their users' information diet and increase the profitability of items in their diets by decreasing the amount of energy expended when foraging for desirable items. Allowing users to take advantage of the paths created by others through collaborative filtering, as Wexelblat (1999) demonstrated, leads to greater user satisfaction and greatly reduces the cost associated with foraging.
Card, Stuart K., Peter Pirolli, Mija Van Der Wege, Julie B. Morrison, Robert W. Reeder, Pamela K. Schraedley, Jenea Boshart (2001). Information scent as a driver of Web behavior graphs. Proceedings of the Conference on Human factors in computing systems CHI '01 Association for Computing Machinery.
Chi, Ed H. Peter Pirolli, Kim Chen, James Pitkow (2001). Using Information Scent to Model User Information Needs and Actions on the Web. In Proc. of ACM CHI 2001 Conference on Human Factors in Computing Systems, pp. 490-497. ACM Press, April 2001. Seattle, WA.
Furnas, G. W, (1997). Effective view navigation. In Proceedings of the Human Factors in Computing Systems, CHI '97 (pp. 367-374). Atlanta, GA: Association for Computing Machinery.
Heer, Jeffrey Ed H. Chi (2000) Identification of Web User Traffic Composition using Multi-Modal Clustering and Information Scent. in Proc. of the Workshop on Web Mining, SIAM Conference on Data Mining, April 2001, Chicago, IL. pp. 51-58
Krug, Steve (2000). Don't make me think: a common sense approach to web usability. Macmillan USA.
Marr, D. (1982) Vision. San Francisco: W.H. Freedman.
Pirolli, Peter, Stuart Card (1995). Information Foraging in Information Access Environments. In Proceedings of the Human Factors in Computing Systems, CHI '95. Association for Computing Machinery.
Pirolli, P. (1997). Computational models of information scent-following in a very large browsable text collection. In Proceedings of the Conference on Human Factors in Computing Systems, CHI '97 (pp. 3-10). Atlanta, GA: Association for Computing Machinery.
Pirolli, Peter, Stuart Card (1999). Information Foraging. Psychology Review Vol. 106, No. 4. (pp.643-675)
Reisberg, Daniel (2001). Cognition: exploring the science of the mind. 2nd ed. W.W. Norton & Company, Inc.
Russell, Daniel M., Mark J. Stefik, Peter Pirolli, Stuart K. Card (1993). The cost structure of sensemaking. Proceedings of the Conference on Human Factors in Computing Systems. Association for Computing Machinery.
Ward, David, Jacob Blaustein (1992) The role of satisficing in foraging theory. Oikos 63:2 (pp. 312-317).
Wexelblat, Alan, Pattie Maes (1999) In Proceedings of the Human Factors in Computing Systems, CHI '99. Association for Computing Machinery.
posted by Dr. Adam Norten at 9:13 AM
0 Comments:
Post a Comment
<< Home