This a technical problem discussed in Social Media Analysis 10-802 in Spring 2010.
Given a resource (link, photo etc.) and tags assigned to by the user, find tag to enrich this annotation.
- Present them to user, who can select them to add them to the resource
- Add them automatically to the index entry of this resource to increase search performance.
A non exhaustive list of some methods that have been used to approach this problem.
Find related tags by looking which tags co-occur often.
Represent tags as vectors, where the indices in the vector indicate how much the tag co-occur with other tags/resources/users. Then using similarity measures such as Cosine similarity related tags can be found.
Latent dirichlet allocation for tag recommendation, Conference On Recommender Systems, 2009 uses LDA for tag recommendation. They build a topic model on training data. For a new resource, they recommend the top tags of the inferred topics for that resource.
Methods build on Folkrank (modified PageRank algorithm for folksonomies).
The main challenges with this problem is to deal with very sparse data. Furthermore, to make it practical for real systems, it also needs to be very scalable.
Want to know more?
- ECML PKDD Discovery Challenge 2009 had three tag recommendation tasks (content based, graph based and a task where performance was measured in an online setting)