Overview-Based Example Selection in End-User Interactive Concept Learning

Comments:
None Yet

Summary:
This paper aimed to demonstrate a new program called CueFlik.  CueFlix's goal is to create a learning agent that will allow users to better access large datasets.  The paper describes the problem's users face when accessing large datasets using key words, explaining how the nature of keyword search forces users to have to go through large amounts of data manually to find what they are looking for.

CueFlix attempts to fix this problem by learning about what the user is looking for.  This program focuses on image search and trying to narrow down the results from a keyword search on images.  To do this the user will input a keyword such as "radio" and a grouping of images will be displayed.  Then the user will select images similar to what they are looking for.  For example if they are looking for a cartoon image of a radio the user would select and submit the cartoon images.  This will tell CueFlik to display more cartoonish images.

Discussion:
The paper outlined the methods used to help narrow down searches and reduce the amount of extraneous results given to the user.  This is significant because it allows the user to better tailor results to them instead of having to wade through the results that the search engine thinks the average user would most likely enjoy.

The biggest problem with the work was stated in the user comment, "it was weird, sometimes it would start out doing really well, but as I kept going it did worse” and “it is hard to know if more data is better as I should probably stop occasionally to see the results as I am going."  The general problem with learning systems such as Pandora, or this one is that the user doesn't know when to stop training the system.  The user has anxiety about over training the system and then pigeon holing themselves into a small niche that they might not even know they are in but they don't want to under train the system and not get the full range of functionality that they could be getting.

If I were in charge I would definitely keep the project going.  It seems to have a lot of promise but there is some refinement that needs to be done.  For example the previously mentioned problem with letting the user know when to stop teaching the system needs to be addressed more in depth than is already being addressed.