Learning to Generalize for Complex Selection Tasks

Comment on Ross's blog.

Have you ever been going through a folder on your computer trying to select only certain files. It can be really frustrating and tedious. One misstep can deselect all of your carefully made selections. Alan Ritter from the University of Washington and Sumit Basu from Microsoft Research have been looking for a way to make complex selection easier.



The first component of their selection tool is the selection classifier. This component looks for features common to the selections the users are making and weights them. Some example features include:

• The presence of any substrings of length 3 or greater in the files' names
• The value of the file extension
• the file creation date being greater or less than/equal to each of the file creation dates of the current selections
• the file size being greater or less than/equal to each of the file sizes of the current selections.

Next the regressor tries to generalize the users selections so that it can accurately predict what other files the user is going to select. Some of the features that the regressor looked at for item i were:

• The number of times the user (de)selected an item while i was (de)selected
• Whether the last example provided by the user last changed the selection state in the same round as i
• The proximity of i to the last example provided by the user.
• and more

For the user study, participants were asked to select certain groups of files using regular selection and their new auto selection tool. The tool was found to be highly accurate and many of the users said that they enjoyed it.

I would love to have this feature on my computer. This would make reorganizing files on my computer a lot easier and quicker. My only fear is that it might auto select items that I want to keep when I am selecting things I want to delete. I would be really paranoid about this and probably double and triple check.