Reorganizing clouds: A study on tag clustering and evaluation

Abstract

Finding and visualizing semantic relations among tags within a tag cloud enhances user experience, particularly regarding access to and retrieval of web pages on social tagging systems. Several approaches have been proposed to visualize tag relations in these systems. However, results of previous research rely on qualitative evaluation methods, and do not provide robust and sound comparison criteria. In order to allow quantitative evaluation we present a benchmark social tagging dataset, where a subset of 140 tags from a well-known social bookmarking site, Delicious, have been manually categorized according to the Open Directory Project (ODP). The manual categorization is utilized as a ground truth that enables quantitative evaluation providing a way of inferring the best of different clustering approaches. With this dataset we also explore different tag representation approaches to present a reorganized tag cloud by using Self Organizing Maps. In addition, we present an approach to enrich the resultant tag cloud with the most characteristic terms for each tag and group of tags, making possible a further filtered navigation, both by tag and document content, and easing a deeper qualitative evaluation of the clusters.

A study on human-generated tag structures to inform tag cloud layout

“…Approaches that propose automatic algorithms for generating structured tag clouds include [7,1,16,9,6,13,5]. Most of them approximate what they call the underlying "semantics" or "meaningful relations" between the tags with a cooccurrence-based measure.…”

“…Regarding the layout, i.e., the visual representation of the determined term relations, many different variants have been proposed, e.g., line-wise representations with one line per cluster [7] or with 2D clusters whose terms are spread across multiple lines [18]; network representations [9,6]; tag hierarchies [5]; 2D arrangements in which closeness reflects the relatedness of terms [21,4,15], some of which emphasize clustering structures [3,1]; and SOM-based visualizations [16]. Only few of these suggested visual representations were evaluated with respect to their suitability and appropriateness for the specific task.…”

Automatic content extraction and visualization of Thai websites for improved information representation

“…It is a means to represent the characteristics of the original content [2] and this is a form of visualization or summary [20,21] as they provide conceptual information of the original content [22]. The importance of the tags could be visualized based on size, color and style.…”