Database Reference
In-Depth Information
a power law [ 32 ], and the probability of a tag having tag frequency x is proportional
to x 1.15 . The head of the power law fit contains tags that would be too generic to be
useful as a suggestion, while the tail contains the infrequent tags that often corre-
spond either to wrong words or to highly specific tags. Furthermore, the distribution
of the number of tags per photo also follows a power law distribution. The
probability of having x tags per photo is proportional to x 0.33 . The head of the
power law contains photos annotated with more than 50 tags and the tail contains
more than 15 million photos with only a single tag, while almost 17 million photos
have only two to three tags. The majority of photos is thus annotated with only a few
tags, which describe where the photo is taken, who or what appears in the photo,
and when the photo was taken.
Different works addressed the research issues on photo annotation. To support
automatic photo annotations, diverse tag recommendation techniques have been
studied. As shown in the taxonomy, depicted in Fig. 2.3 , the proposed approaches
can be classified into basic techniques and collective knowledge supporting.
Flickr offers a service to suggest tags when a user wants to tag a picture.
Suggested tags, sorted lexicographically, include recently used tags and those
most frequently employed by the user in the past. However, this service is rather
limited. One step further toward personalized tag suggestion for Flickr was pre-
sented in [ 27 ]. Three algorithms have been proposed to suggest a ranked list of
tags to the user. Proposed algorithms receive as input parameters the identity of the
user, an initial set of tags (if available), and the corresponding tagging history of all
users. The recommendation is based on the tags that the user or other people have
exploited in the past. Furthermore, the suggested tags are dynamically updated with
every additional tag entered by the user. The first two methods consider only those
tags exploited by the user in the past and suggest a ranked list of tags, sorted by
considering both tag frequency and past inserted tags. The last method considers
both the set of tags exploited by other people in the past and the tags similar to those
entered by the user for the same picture in the past. In particular, a set of promising
groups is first identified by analyzing both the user and the group profiles. Then, for
each of these groups, a ranked list of suggested tags is generated according to tag
frequency and past inserted tags.
To validate the methods proposed in [ 27 ], different pictures have been down-
loaded and divided into two groups: (a) 200 pictures with four to eight given tags
and (b) 200 pictures with more than ten given tags. The method which considers
tags exploited also by other people is more effective in suggesting relevant tags.
Furthermore, the results obtained for the second set of pictures are better than those
obtained for the first set because users who add more tags to an individual picture
usually have a better tagging history. In fact, the methods proposed in [ 27 ] yielded
better accuracy on pictures with a large number of given tags.
Different and more effective approaches have been proposed to effectively
support photo annotations [ 27 , 31 ].
Authors in [ 31 ] proposed to exploit the collective knowledge that resides in the
Flickr community to support tag recommendation. Given a photo, the proposed
recommendation system selects a list of relevant tags. The proposed system operates
Search WWH ::




Custom Search