upload old file or otherwise resume clustering after shutting down setting

As a follow-up from #1, when clusters are updated, the titles should update accordingly. Commit a9e13ae sends the new titles to the HTMLwidget, but the text itself doesn't update yet. @gwgundersen, could I get some help finishing this off?

Likely related to Issue #28

We should find the limits on size of data and value of K for which the tool runs smoothly, and work on supporting larger data sets and higher values of K

use only vocabulary; search with stem and unstemmed

Shift+click for clusters should move the cluster as a whole, instead of creating a new cluster containing the cluster (without the shift, it merges clusters). Alternatively, we could add a different key modifier and keep the current behavior. Regardless, we need a way to move clusters.

@AFriendlyRobot want to tackle this one?
@gwgundersen advice welcome

When double clicking to select a node, other text visible on the screen can become selected/highlighted. This is not necessarily the closest text.

Currently, zooming does not change which topic titles are displayed in the bubble widget, no matter how deep the hierarchy is. Zooming should allow more titles to be displayed (hierarchically).

The current system uses a single value (k) to denote a "zooming" ratio. However, no matter how large the window is at initialization, k is initialized to 1. This particularly causes text resizing issues when start in a small window and increasing the size vs. starting in a small window and shrinking the size.

Model output: https://drive.google.com/file/d/0B8BAFw2W1pO1OFphTkFKcG1ESGs/view?usp=sharing

This seems to happen fairly frequently, but I have not yet identified exact conditions necessary or sufficient to produce the behavior. Reported error:

Note that the numbers at the end seem to change, and I have been able to produce this error in several trials by moving nodes around and then increasing or decreasing the number of clusters parameter.

Currently, we statically load the top-weighted 100 documents per topic. Ideally we will be able to dynamically load more documents as the user scrolls. This will increase performance while also allowing the user to explore potentially all documents.

See discussion of #1.

When changing between the bubble widget and tree widget, some visual inconsistencies can occur. For example, if the user selects node A in the bubble widget, then switches to the tree widget and selects node B, then switches back to the bubble widget, node A will still APPEAR selected. However, the backend will accurately consider node B to be selected (displaying documents related to node B on the sidebar).

One simple option is to clear selected nodes when changing between widgets, but this is not necessarily the best behavior.

Create a smaller data set (than the wikipedia data set) which can be uploaded to the repository.
This should follow the current format of the .RData file, and will allow for faster iteration during development and debugging.

If I zoom into a cluster and create new clusters within that cluster, or rearrange leaves within that cluster, I don't want to zoom out every time I make an edit.

@gwgundersen thoughts on this welcome
@AFriendlyRobot might be another good issue to start with

video of current interaction: https://www.dropbox.com/s/qte3p6b6hjfa9y1/zoom_repeat.mov?dl=0

At some point, we may want to allow document metadata to be displayed for each document (if provided).

One solution: In the directory of text files, also include RData files containing whatever data should be displayed.

If a node is collapsed before a drag-and-drop merge of clusters (in the tree view), the underlying data structure will correctly merge/move. However, visually, the previously collapsed/hidden nodes will remain collapsed/hidden but the newly moved nodes can appear as children of a collapsed node.

Suggested fix: Uncollapse nodes when merging (seems like a reasonable behavior)

Currently exporting an SVG image does not include the associated CSS styling.
Solution: Create a dummy copy of the SVG elements, using Window.getComputedStyle() to copy style explicitly.

@shiraito noticed this using Chrome

per discussion in #1

Desired:

1. Expand volumes available
2. Improve file/directory path parsing

Potential solutions

1. Use shinyFiles getVolumes() augmented with custom code (primarily for Windows)
2. Use shinyFiles parseFilePath()/parseDirPath() methods

after #1

Clustering within clusters.

Deleting characters of the edited title will update the displayed title, except for the last title.
Similarly, selecting edited title and deleting at once will not update the title.

After moving node C (with edited title) from parent P1 to parent P2, on first mouseover of C, C's old title will display. Then new (edited) title will display, while all visible titles redisplay. Then C's title will go blank

In the panel that displays the contents of a document, it would be good to somehow display the weights of all topics for that document.

When a leaf node is selected, it must either be moved to a new parent or a different leaf node must be selected. We should be able to unselect it.

@gwgundersen not sure if this is a side-effect of the hierarchical changes or not (thought we used to be able to do this).

Do we want to have leaf nodes able to exist in the root node without being assigned to a cluster? My gut is no. @bstewart @shiraito, confirm?

@AFriendlyRobot want to tackle this too?

An evolution of #4

Per discussion for #1, we want to make sure the users' attention is drawn to the right things.

Also ensure that colors adjust depending on the depth of the hierarchy.

When hovering over a topic, a Topic Summary panel appears in the left column. However, if the user scrolls while this panel is open in certain ways (see below), the panel will hide itself and then reappear, causing the page to expand and contract and causing the visualization to appear to bounce up and down.

Conditions for this to occur:
It appears that when the user (starting fully scrolled to the top of the page) hovers over a topic bubble and opens the panel, then scrolls down so that the mouse is hovering over the root node only, the topic summary panel will hide (no specific topic with a summary is hovered). However, this shrinks the total height of the page, which effectively pushes the mouse back up. If the positioning of the mouse is "correct", this new position will be over a topic bubble.
The topic summary panel will then expand again. Because the page was scrolled to the bottom previously, Chrome at least (not yet tested on other browsers) will reset the scrolled position of the page to the bottom. Therefore, the mouse is no longer hovering the relevant topic bubble, and the cycle continues.

Likely requires a little more examination to fully identify all necessary and sufficient conditions.

@ajbc Again, I'm not sure if this is necessary to address before Text as Data. However, I don't think any overhaul we've discussed would be likely to address this issue specifically, and this issue is therefore probably independent from much of the potential work for the Overhaul milestone.

Allow users to color-code or flag individual nodes/topics/clusters. E.g. a user could change a cluster to green if they think the cluster is well-formed.