Paper Reading #25: Tagsplanations: explaining recommendations using tags

Comments
Evin Schuchardt
Luke Roberts

Reference Information
Title: Tagsplanations: explaining recommendations using tags
Authors: Jesse Vig, Shilad Sen, John Riedl
Presentation Venue: IUI 2009: Proceedings of the 14th international conference on Intelligent user interfaces; February 8-11, 2009; Sanibel Island, Florida, USA

Summary
This paper discusses tagsplanations, explanations based on community tags that are made up of two parts: tag relevance (relationship of the tag to the recommended item) and tag preference (relationship of the user to the tag). They discuss an algorithm used to determine tag relevance and tag preference and present a study to see how users like tagsplanations.

Image taken from paper: Results sorted by tag relevance

To test the implementation of tagsplanations, the researchers use MovieLens, a movie recommendation site. They compute tag preference based on user behavior (by movie ratings in their MovieLens implementation). For example, to determine a user’s preference for the tag, violence, they would look at the user’s ratings of other films with the tag, violence, and calculate a weighted average for the preference value. They use the weighted average because some movies have been tagged with a certain identifier more than others and thus should be weighted more during the calculation. For tag relevance, they calculate the correlation between a user’s preference for a tag and their preference for the movie.

In their study the researchers sorted results in four different orders and then checked to see how well each interface helped users understand why an item was recommended and decide if they would like the recommended item. The researchers also checked to see if the recommended item matched their mood.

The results of the study showed that tag preference was more important than tag relevance but that tag relevance seemed to be the best way in which to organize the results. The interface that only used tag relevance (ordered by tag relevance and displaying tag relevance only) seemed to work the best for mood compatibility. They also found that subjective tags (tags expressing user opinions) performed better than factual tags (tags identifying facts about the movie such as concepts, people or places) overall.

Discussion
While I have found current recommendation systems to work just fine for me, I do think the researchers have presented some good ideas here that can make the entire system more informative and perhaps more relevant. In looking at their interface, I think they’ve done a good job of condensing a lot of information into a small amount of space.

Overall this was an interesting paper. They explained their motivations well and provided an in-depth user study. For future works the researchers mention making it so that users can inform the system when it is wrong and the system can inform users as to how different actions impact the results they receive. They also mention exploring other techniques to estimate tag preference.