If you’re familiar with the mechanism by which Pandora recommends songs, you’ll know that the songs are scored according to certain attributes that all songs have, like Major/Minor key tonality, presence of harmony, etc. This is different in two major ways from how tagging usually works. First, various items of content can have various tags, from a fairly unrestricted vocabulary. Second, the tags are usually user-generated. With Pandora, the attributes of the songs are annotated by musical experts, working for Pandora, using a controlled vocabulary. These differences help explain why Pandora works well, whereas a folksonomy approach only sorta works.

Having spent quite some time trying and testing social media tools for scientists, I’ve become even more jaded than Mr. Crotty himself about their potential. Anything that requires someone to put in even the minimal effort required to tag articles will probably see relatively low uptake. In contrast, a Pandora-like approach would allow a service implementing it to offer value to the user immediately, instead of only after building a large collection/tagging many articles/making a large number of contacts/etc. To get good recommendations from Pandora, all you do is enter a song or artist and then sit back and listen. Likewise, a corpus of papers which had been expertly scored according to a set of attributes such as “uses the technique of …”, “written by a student of …”, and so on could power a recommendation engine requiring little user input other than interesting/not interesting.