Sigurbjörnsson, WWW2008

From Cohen Courses
Revision as of 11:42, 3 September 2010 by WikiAdmin (talk | contribs) (1 revision)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigationJump to search

Citation

Borkur Sigurbjornsson and Roelof van Zwol. Flickr tag recommendation based on collective knowledge. In WWW ’08: Proc. the 17th international conference on World Wide Web, pages 327–336, New York, NY, USA, 2008. ACM.

Online version

WWW2008 website

Abstract from paper

Online photo services such as Flickr and Zooomr allow users to share their photos with family, friends, and the online community at large. An important facet of these services is that users manually annotate their photos using so called tags, which describe the contents of the photo or provide additional contextual and semantical information. In this paper we investigate how we can assist users in the tagging phase. The contribution of our research is twofold. We analyse a representative snapshot of Flickr and present the results by means of a tag characterisation focussing on how users tags photos and what information is contained in the tagging. Based on this analysis, we present and evaluate tag recommendation strategies to support the user in the photo annotation task by recommending a set of tags that can be added to the photo. The results of the empirical evaluation show that we can effectively recommend relevant tags for a variety of photos with different levels of exhaustiveness of original tagging.

Summary

Tagging systems have become very popular the last couple of years. Such systems allow users to upload or link to resources, and annotate them with tags. The paper performs an analysis on Flickr and proposes and evaluates different tag recommendation strategies.

Dataset

Their dataset is a snapshot of Flickr, which contained 52 million photos uploaded between February 2004 and June 2007. Each photo had at least one tag.

Tagging analysis

They show that the tag frequency follows a power law. The head of this power law contains tags that are probably too generic to be recommended, while tags that occur at the tail are very infrequent and often misspellings, complex phrases or very incidentally occurring tags. Thus in most cases tags in the middle of this power law are most useful to recommend to a user. The number of tags per photo also follows a power law. They show that a large proportion of the photos (64%) has only 1 to 3 tags. For these photo's tag recommendation can be very useful. They also mapped the tags to WordNet. This showed that the users annotate tags in a wide spectrum, not only describing the visual contents but also the broader context (such as location, kind of camera etc.)

Tag recommendation strategies

For every tag that was already assigned to that photo, using co-occurrence metrics a list of candidate tags was created. Then using an aggregation step and promotion step, these lists are merged and a final ranking is created.

Experimental setup

Given a photo and the tags already assigned to by the user, present a ranked list of relevant tags. Evaluation was done using relevance assessments by human assessors. The following metrics were used for evaluation Mean Reciprocal Rank, Succes at k (k=1 and k=5) and precision at k (k=5).

Results

When comparing the aggregation strategies, the strategy which uses the co-occurrence values is more effective. Furthermore, the promotion step further improves results. Their best strategy had a relevant tag on the first position in 67% of the cases and 94% when considering the first five tags.

Related papers