The system organizes the pages into clusters based on closeness of pages in various attribute spaces. The interface must take that big ball of connected stuff and turn it into something a user can easily organize, reorganize, navigate, search, and remember. Since most people can't see more than two or three dimensions of variation simultaneously, not all pages and their relationships can be displayed at the same time. Hence much of this information must first be clipped before display. There are three natural clipping stages:
Stage I
The first stage is to clip the data down to some set of pages with some set of distinguishing attributes. These distinguishing attributes might be chosen on the basis of:
variance of attribute values (that is, we might want the distinguishing attributes to be the ones that most discriminate amongst the pages we must display),
Certain attributes are likely to be included in any such clip operation. For example, any attribute indicating past user interest in a page is likely to be useful; any attribute reflecting page content is likely to be indispensable; any attribute reflecting time of creation (or last viewing time, or the length of the last view time, and so on) is likely to be useful. Conversely, attributes involving page size are likely to be less useful---but not irrelevant, so they can't simply be ignored during every clipping operation.
We might also clip not just the attributes, but also the pages. That is, we might choose not to display pages that are rarely looked at whether or not they have attributes in the clipped set of attributes. Or in one particular session we might choose to focus only on a neighborhood of pages rather than the entire page space.
Once the distinguishing attributes are chosen, the pages are then clipped to only those pages that have those attributes, then those pages are clipped to the subset of pages that have the appropriate values of those attributes.
Stage II
The second stage is to map the chosen set of distinguishing attributes to visual features. The single most important visual feature is location in a space, but we can also have color, animation sequence (if any), intensity, transparency, Z-order, sound, and appearance. "Appearance" might be whether it's a cartoon face, an abstract three-dimensional object with shading and texture, a two-dimensional colored icon, or something else.
It's likely that some attributes will be more important than others and it's also likely that some visual features will be more important than others. It seems reasonable to map the more important attributes (content, say) to the more important features (location, say).
It probably will be useful to have visual features that apply to multiple pages, not just each single page. The user could then assign, say, an image of a clef note to all the music pages, or an image of a computer to all the system-related pages, and so on. To aid this process, the system might maintain a collection of clip-art and the user can choose which graphic goes with which category. All the user then sees is the graphic and that serves as a reminder of what's there.
Stage III
The third stage is to take the data and map it into a viewport (in other words, clip the data based on its location relative to the window the user can actually see through at any one time) and handle interaction with the user (aiding navigation and updating the user's current context mainly).
There might be multiple clippers and multiple mappers and multiple viewports. The user might be allowed to choose between them either implicitly by which ones get chosen most often or explicitly for special occasions. However:
So we have two conflicting goals: how to manage an interface that must be both constant and changeable at the same time.
