In this paper, George Furnas provides a framework for discussing the issues relating to design of information spaces for interactive exploration. From this framework, requirements for effective navigation are developed, several strategies for designing easily traversed structures are described, and the requirements for information design in a variety of circumstances are described.
This paper defines two distinct categories of user movement through an information space. View traversal is defined as being an interactive process of viewing a space, selecting an object, and moving to the chosen object. View navigation is traversal with the intent of making good decisions towards the goal of identifying a desired target. These cases are considered separately in the paper, but in the end the goal is the same for both: the identification of strategies for laying out views in order to support effective exploration of large structures.
The discussion of traversability assumes information structured in terms of a logical structure graph, and a viewing graph. The logical structure graph is a graph representation that connects elements "to their logical neighbors as dictated by the semantics of the domain", while the viewing graph is a "window" on the logical structure graph: the subset that is accessible at any given point during the user's travels through the graph.
Effective View Traversability is defined in terms of two requirements: small view and short paths. Requirement EVT1 (small views) demands that the number of out-going links of nodes in the viewing graph must be "small" compared to the size of the structure, while requirement EVT2 (short paths) states that the distance between pairs of nodes in the viewing graph must be "small" compared to the size of the structure. In other words, the number of choices to be made from any node must be minimized, and the number of steps needed to get from any one node to any other must also be minimized.
>From the EVT requirements, several uncontroversial conclusions can be drawn: straight lines are not efficiently traversable, while balanced trees and hypercubes are. Structures that are not EVT - such as straight lines - can be fixed by folding the line, using fish-eye sampling, or adding a tree structure. Furnas claims that these strategies can be generalized: "always remember the strategy of putting a traversable infrastructure on an otherwise unruly information structure!"
Navigability is defined in terms of the assumption that each node in a structure has information on each outlink (outlink-info) that indicates where the link will take the user. The goal is strong navigability: the structure and info should combine to help users find the shortest path to the target, without error, and based only on local information. Several definitions regarding the sets of nodes that can be reached from links and the quality of outlink information are presented, with the result being requirement VN1 (navigability): "the outlink-info must be everywhere well-matched". Informally, the outlink information available at each node must be complete and accurate.
Seen from another viewpoint - that of the target - each target must present information to incoming links, indicating that the target can be found at the end of the link. This information is known as residue or scent - just as animals leave scent for predators, nodes in a logical structure graph may leave scent for information seekers. This leads to a restatement of the first requirement, as requirement VN1A (residue distribution): "Every node must have good residue at every other node". Unfortunately, this is in direct conflict with requirement VN2: "outlink-info must be "small". In other words, the two requirements (VN1a and VN2) combine to require that each node must tell every other node how it can be reached, but this must be done such that no link has an outlink-info set that is significantly large with respect to the entire structure: this simply does not scale well.
>From these requirements, Furnas concludes that strong navigability is not possible, and that outlink information must be labeled in some way that provides appropriate characterizations. In other words, residue must be shared by many nodes, so that the outlink information available at a node can succinctly describe all of the nodes that are found. For example, while it may not be possible to label a link in a graph of animals with all of the species that may be found by following that link, an outlink label such as "cats" allows all of the nodes that can be found from that link to share the residue in a space-efficient manner. Similarity-based navigation may be useful, but it runs into the "local maxima" problem inherent in hill-climbing structure: since only local information is used, the navigator can never be certain that they are progressing towards the goal.
Furnas concludes with some remarks on combining Effective View Traversability with View Navigation. First, "Large scale semantics dominate" - each set of outlink information must account for large number of nodes, or else VN2 will be violated. Furthermore, effective view navigability is contingent upon an efficient semantic decomposition of the information, with minimal overlap of outlink information.
Unfortunately, some of the arguments and definitions were not entirely clear, and almost counter-intuitive. For example, EVT1 (small views) seems like more of a constraint than a requirement. Given the extensive research arguing for a preference for broad and shallow menu trees (which would fit will with EVT2), it would seem that designers should strive for including as many nodes as possible in a given view.
The discussion of traversability also seems to ignore semantic issues: the comments on fixing non-EVT structures do not address the semantic impact of changes and additions to the logical view graph. Perhaps Furnas is assuming that any such changes are unimportant, as traversability does not imply looking for a known target. However, the utility of techniques such as folding of lists or super-imposing a tree structure would seem to be highly dependent on the type of information in the structure, the topological effects of any added links, and the labels of added nodes. For example, it seems to me that the ordered version of the list of spices given in figure 1 would be much more traversable than the folded version given in figure 3c. Even if traversal of the ordered list requires more steps, the information provided by the alphabetic ordering is quite useful, and may be completely lost in the folded version.
While I found the description of navigability (and the associated figure) to be very unclear, the model of strong navigability and the two versions of requirement VN1 seemed to be useful conceptual points.
Some of the terminology found in this paper raised some questions. Specifically, the terms "residue" and "scent" do not seem to hold up particularly well. While the idea of presenting hints that say "This piece of information can be found by following this link,&quoyt; seems appealing, residue or usually apply to the impact of the actual presence of that which leaves the residue or scent. In the information structure case, use of this word would seem to imply that the information found at a given node had already "been at" the node where the residue has been left. In some sense, the model seems to break down in this case.
However, the description of navigation through a biological taxonomy pointed out one of my main issues with this paper. Furnas uses a discussion of scent and residue to claim that the efficient navigation of these taxonomies is not as trivial a matter as it may seem. However, he fails to acknowledge the real reason why this particular hierarchy works so well: our fairly-well developed understanding of the domain involved. The residue is not an inherent property of the graph, but is in fact intimately tied to the user's domain knowledge: without this knowledge, the structure would most likely be useless.
This example seemed to be an illustration of the main problem with this paper: by addressing the questions of traversability and navigability strictly in terms of the layout of the information structure, Furnas fails to directly addresses the impact of factors regarding the type of information involved, and our cognitive models and domain knowledge. The impact that these factors have upon traversability and navigability might, in many cases, be as significant as the information structure itself.