Improving Tag-Clouds as Visual Information Retrieval Interfaces

• Tagging uses freely chosen keywords to categorize resources for later refinding, and shared tags form a folksonomy that makes tagged resources accessible.Source: §1 · p. 1
• Folksonomy access includes passive filtering through subscriptions to tags and active retrieval through tag queries, which can lead to browsing co-occurring related tags.Source: §1 · p. 1
• A Tag-Cloud displays popular tags alphabetically with font-size weighting, enabling users to scan a global, contextual view and recognize information needs without first querying.Source: §1 · p. 1, Figure 1 · p. 2
• Frequency-only selection yields semantically dense clouds, while alphabetical layouts hinder scanning and fail to convey relations; the paper argues for improved selection and similarity-based layout.Source: §1 · p. 1
• Co-occurrence analysis can reveal associative and hierarchical tag similarities, supporting similarity-based layouts that convey relations among tags.Source: §1 · pp. 1–2
• The proposed method reduces semantic density in the selected tag set and groups tags by similarity using clustering techniques and co-occurrence analysis.Source: §1 · p. 2

2 BACKGROUND

• Tags may be polysemous, synonymous, or ego-oriented, and often identify authors, opinions, or resource types rather than topics [26][3].Source: §2 · p. 2
• Folksonomies reflect users’ vocabulary [15], derive meaning through collective use, reduce inter-indexer inconsistency [7], and support serendipitous discovery [15].They can also provide user-centered indexing where controlled vocabularies are impractical.Source: §2 · p. 2
• Specificity counts resources described per tag, while exhaustivity counts tags per resource; broader tags favor discovery and browsing, whereas narrower tags favor recovery and querying.Broader tags imply higher recall and lower precision; narrower tags imply lower recall and higher precision.Source: §2 · pp. 2–3
• Taggers tend to assign broad tags, and at least 90% of cases involve fewer than five tags per resource; their low specificity makes tags better suited to browsing than querying.The section frames this as an answer to whether folksonomies better support browsing or querying, and relates low cognitive effort to tagging’s popularity.Source: §2 · p. 3
• Prior visualizations include a time-evolution view of tags, a graph-based tag browser, and a circular cloud encoding tag importance but not inter-tag similarity.The section notes that few such visualization works target information retrieval.Source: §2 · p. 3

3.1 Sample

• The October 2005 del.icio.us sample contains 218,063 URLs, 242,349 tags, and 111,234 users, collected with an ad-hoc crawler.The system allowed any user to tag any resource and hid other tags on a resource during tagging.Source: §3.1 · p. 3

3.2 Tag Similarity

• Tag similarity is defined as relative co-occurrence: the number of resources bearing both tags divided by the number bearing either tag.The paper also calls this co-occurrence-based similarity tag overlapping.Source: §3.2 · p. 3

3.3. Tag Selection

• Tag usefulness is designed to select tags that characterize the whole collection, balancing resource representation, coverage, and distinctiveness.Its simplified length-normalized tf·idf-based function uses log(d_ij) to reduce the effect of frequent tags within a resource and m^2 to downweight less discriminative tags; the numerator captures representation and the denominator discrimination value.Source: §3.3 · p. 4
• Approach (d) reduces tag overlap versus traditional methods (a, b) and the same function without m^2 (c), while coverage differences are minimal.The comparison evaluates selections of the 95 highest-weighted tags; coverage is the number of resources described by at least one selected tag.Source: §3.3 · p. 4, Table 1 · p. 4

3.4. Tag-Cloud Layout

• The layout clusters the 95 highest-weighted tags into 12 groups using bisecting K-means on their relative-co-occurrence similarity matrix.Cosine similarity measures co-occurrence patterns; tags are arranged horizontally beside similar tags, and similar clusters vertically beside one another.Source: §3.4 · pp. 4–5

4 RESULTS

• The improved Tag-Cloud shares 65 tags with the traditional cloud while removing less significant, site-specific, and synonymous tags.Examples removed include ‘toread’, ‘del.icio.us’, and ‘webdev’.Source: Figure 2 · p. 5, §4 · p. 5
• More discriminative topics emerge and reduce semantic density, although tech topics remain dominant because of del.icio.us’ tech-savvy user base.Examples of emerging tags include ‘philosophy’, ‘health’, ‘food’, ‘fashion’, and ‘cinema’.Source: §4 · p. 5
• Clustering distributes tags more coherently than alphabetical ordering, distinguishing main topics and enabling users to infer semantic relationships from neighboring tags.For example, ‘ajax’ may appear conceptually related to ‘javascript’ and ‘xml’.Source: §4 · p. 5

5 DISCUSSION

• The paper presents a simple, low-computational-cost method to select the N most relevant tags from a collaborative tag index and display them using a clustering-based layout.Source: §5 · p. 5
• Synonym and plural-form problems remain; the authors suggest a predefined similarity threshold for identifying synonyms because synonymous tags frequently co-occur on the same resource.The paper notes that very similar tags may appear together because of their high co-occurrence frequency.Source: §5 · p. 5
• The approach could support hierarchical navigation through sub-Tag-Clouds showing the N most relevant tags on resources associated with a clicked tag, excluding that tag.Source: §5 · p. 5
• Future work aims to improve tag selection and weighting and develop alternative approaches and graphic metaphors for visualizing large folksonomies.Source: §5 · p. 5