Guided Research Bernhard Waltl

Many web sites use tags, i.e., freely-chosen text labels, as a means for ca-tegorizing contents created and uploaded by independent users (e.g. photos,bookmarks, music, podcasts) in order to make these contents accessible. Wherehierarchical organization schemes fail, tags offer a way to navigate, search andbrowse through these collections of information resources. The entire set of tagassignments in such a collection is called a folksonomy. These rather chaotic organization schemes exhibit no clear structure since there are no constraints inthe usage of tags and users have different vocabularies. In combination with thefact that the number of resources can be very high this means that the utilityof the system sffers if the user is not provided additional aid in analyzing a setof resources. A common approach is to analyze the number of co-occurrencesof tags and generate clusters of tags being frequently used together. Severalalgorithms for this purpose are described in literature, sometimes includingspecific user interfaces displaying the results to the user.

In this work, an alternative approach shall be systematically studied. Basedon the co-occurrences of tags as well, an algorithms shall be developed thatpartitions the collection of resources by finding sets of pairwise mutuallyexclusive tags. Each partition corresponds to the resources tagged with one tagin such a set. This way, it is expected to be able to reliably detect tags thatbelong to the same dimension of categorisation, such as "sad" and "funny"or "red", "green" and "blue". These tags can then be used to offer a morestructured way of accessing and _ltering the collection.

In a first step, a literature study is conducted in order to identify partitioningalgorithms being applicable to the problem. Particularly, the algorithmsshould be able to detect fuzzy partitionings in the sense that partitions mayoverlap to some extent or there is a rest of resources not being containedin any partition. If no such algorithm can be found, it will be determinedin how far existing other partitioning algorithms can be adapted to the problem.

Subsequently, one such algorithm will be implemented and applied to data samp-les from real world folksonomies. Finally, the results of the partitioning will bequalitatively analyzed to assess if meaningful facets can be extracted.