Source-linked AI summary
Improving Tag-Clouds as Visual Information Retrieval Interfaces
Yusef Hassan-Montero, Victor Herrero-Solana
TL;DR
Traditional Tag-Clouds are limited by frequency-based selection and alphabetical layout, which can produce semantically dense collections and weak visual organization. The paper proposes relevance-based tag selection and clustering-based display using tag co-occurrence, reporting reduced semantic density and more coherent visual relationships. It also notes that synonym and plural-form problems remain.
Problem
Traditional Tag-Clouds use frequency-based selection and alphabetical arrangement, producing semantically dense sets and layouts that do not reveal tag similarity.
Method
The paper selects relevant tags using co-occurrence-based similarity and displays them with a clustering-based layout.
Results
The improved cloud removes less significant, site-specific, and synonymous tags, adds more discriminative topic tags, and produces more coherent visual distributions.
Takeaways & Limitations
Clustering-based Tag-Clouds can support browsing by differentiating topics and exposing semantic relationships among neighboring tags.
Takeaways & Limitations
Synonym and plural-form problems remain, although the authors suggest a similarity threshold as a possible solution.
Abstract
from arXiv · showhide
Tagging-based systems enable users to categorize web resources by means of tags (freely chosen keywords), in order to refinding these resources later. Tagging is implicitly also a social indexing process, since users share their tags and resources, constructing a social tag index, so-called folksonomy. At the same time of tagging-based system, has been popularised an interface model for visual information retrieval known as Tag-Cloud. In this model, the most frequently used tags are displayed in alphabetical order. This paper presents a novel approach to Tag-Cloud's tags selection, and proposes the use of clustering algorithms for visual layout, with the aim of improve browsing experience. The results suggest that presented approach reduces the semantic density of tag set, and improves the visual consistency of Tag-Cloud layout.
1 INTRODUCTION
Tagging creates shared folksonomies that support active and visual information retrieval. Traditional Tag-Clouds select frequent tags and arrange them alphabetically, limiting semantic discrimination and visual scanning.
- Tagging lets users categorize web resources with freely chosen keywords, while shared tags form a social folksonomy.
- Folksonomies support information filtering and retrieval, including querying and browsing through related tags.
- Tag-Clouds provide a global, contextual view of popular tags and let users recognize information needs while browsing.
- Frequency-based selection produces semantically dense tag sets, while alphabetical layouts neither facilitate visual scanning nor reveal semantic relations.
- Traditional Tag-Clouds select and visually weight tags by frequency, then display them alphabetically.
- The paper proposes selecting tags differently and grouping them by similarity through clustering and co-occurrence analysis.
2 BACKGROUND
Tagging systems contain broad, ambiguous, and heterogeneous tags, making them more suitable for browsing than precise querying. Prior visualization work motivates improving tag selection and similarity-based layouts for information retrieval.
- Tags may represent polysemy, synonyms, personal intentions, authors, opinions, or resource types rather than topics or subjects.
- Folksonomies reflect users’ vocabulary and collective agreement, potentially providing user-centered and democratic meanings.
- Broader tags offer higher recall but lower precision, whereas narrower tags offer lower recall but higher precision.
- Prior work connects broader tags with discovery and browsing, and narrower tags with recovery and querying.
- Users often assign broad tags and fewer than five tags per resource, making tagging low-specificity and more suitable for browsing than querying.
- Existing visualization studies include evolving tag displays, graph-based tag browsers, and circular clouds, but few target information retrieval.
3 MATERIALS AND METHODS
The study develops tag-selection and layout methods for Tag-Clouds using collection-level usefulness, tag co-occurrence similarity, and clustering. It evaluates selection methods on a large del.icio.us sample and applies bisecting K-means to organize tags visually.
- 3.1 Sample: The study analyzes 218,063 URLs, 242,349 tags, and 111,234 users from del.icio.us, using a crawled sample collected in October 2005.
- 3.2 Tag Similarity: Tag similarity is measured by relative co-occurrence, or the Jaccard coefficient, based on resources shared by two tags.The coefficient divides shared-resource count by the number of resources containing either tag.
- 3.3. Tag Selection: The proposed tag-usefulness function combines representation and discrimination value to select tags characterizing the entire resource collection.Its numerator measures representation, while the denominator gives greater weight to tags describing resources less covered by other tags.
- 3.3. Tag Selection: The proposed selection method (d) achieves better tag-overlap results than traditional methods (a, b) and the unsquared variant (c), while coverage differences remain minimal.Table 1 compares the methods over the 95 highest-weighted tags selected by each function.
- 3.4. Tag Layout: Bisecting K-means clusters the highest-weighted tags from a relative-co-occurrence similarity matrix, using cosine similarity with N=95 tags and K=12 clusters.The layout places similar tags horizontally near one another and similar clusters vertically near one another.
4 RESULTS
The improved Tag-Cloud replaces less significant, site-specific, and synonymous tags with more discriminative topic tags. Clustering produces a more coherent visual distribution that distinguishes topics and supports semantic inferences from neighboring tags.
- The improved Tag-Cloud shares 65 tags with the traditional version and removes less significant, site-specific, and synonymous tags.Examples removed include ‘toread’, ‘diy’, ‘cool’, ‘del.icio.us’, ‘flickr’, and several synonymous tags.
- More discriminative tags such as ‘philosophy’, ‘religion’, ‘health’, and ‘food’ emerge, reducing semantic density while technology topics still dominate.
- Clustering produces a more coherent distribution than alphabetical ordering, differentiating main topics and supporting semantic inferences from neighboring tags.The paper gives ‘ajax’, ‘javascript’, and ‘xml’ as an example of a neighborhood suggesting conceptual relatedness.
5 DISCUSSION
The paper presents a low-cost method for selecting relevant tags and arranging them with clustering-based layouts. It identifies remaining synonym and plural-form problems while suggesting sub-Tag-Clouds for hierarchical navigation and future improvements to selection, weighting, and visualization.
- 5 DISCUSSION: The method selects the N most relevant tags from a collaborative tag index and displays them using a clustering-based layout.The authors describe it as simple and low computational cost.
- 5 DISCUSSION: Synonym and plural-form problems remain, although the authors consider them relatively easy to address with a similarity threshold.High co-occurrence can cause synonymous or semantically similar tags to appear together.
- 5 DISCUSSION: The approach could support hierarchical navigation through sub-Tag-Clouds showing relevant tags assigned to resources under a clicked tag.The clicked tag would be excluded from the new cloud.
- 5 DISCUSSION: Future work targets improved tag selection and weighting methods, along with alternative approaches and graphic metaphors for large folksonomies.
- 5 DISCUSSION: The paper was scheduled to appear at InSciT2006 in Mérida, Spain, from October 25–28, 2006.