Methodologies for Improved Tag Cloud Generation with Clusterin

Size: px

Start display at page:

Download "Methodologies for Improved Tag Cloud Generation with Clusterin"

Vanessa Blake
5 years ago
Views:

1 Methodologies for Improved Tag Cloud Generation with Clustering. Martin Leginus, Peter Dolog, Ricardo Lage, and Frederico Durao Department of Computer Science, Aalborg University July, 2012

2 Agenda Introduction Syntactical pre-clustering of tags Improving coverage and diversity of tag clouds with clustering Experiments

Introduction Tag clouds Problems Methodologies Experiments Conclusion Social collaborative tagging Tagging is a process when a user assigns a tag to an item Users

3 Introduction Tag clouds Problems Methodologies Experiments Conclusion Social collaborative tagging Tagging is a process when a user assigns a tag to an item Users collaboratively annotate items with tags It improves searching, discovering and categorizing of content Tagging information of users (implicit ratings) is utilized by tag-based recommenders

4 Tag clouds A tag cloud is a visual depiction of user-generated tags typically used to describe the content of web sites [Wikipedia, ] Often used as visual information retrieval interface.

5 Common use cases of tag clouds Content Browsing Navigation along the site content Serendipitous discovery

6 Problems The majority of tag clouds are constructed according to the tag popularity/frequency. These clouds suffer from: A cloud does not depict whole spectrum of tag space (missing rarely used tags which can be interesting for users) as only most frequent tags are considered. New tags (new content areas) in a system are hardly depicted in a tag cloud because of they low tag frequency. The most popular tags can be often useless for content discrimination or discovery. (see tags: night, autumn, cool..)

7 Problems Syntactically similar tags cause redudancy

8 Tag cloud s metrics Synthetic metrics express a quality of tags selection process. A coverage for a particular tag t expresses how many of considered documents were annotated with a tag t Coverage(t) = Dt D a, (1) Overlap of T c: Different tags in T c may be assigned with the same item in D Tc. The overlap metric captures the extent of such redundancy. Overlap(T c) = avg ti t j D ti D tj min{ D ti, D tj }, (2) We introduce a new metric chained coverage that captures how many documents are covered by a considered tag given that documents covered by previously selected tags are not considered. This metric combines coverage and overlap altogether Chained coverage(t T s) = Dt \ D T s, (3) D a

9 Syntactical Pre-clustering of Tags Tags in these systems can have the same semantical meaning however they are syntactically different i.e., typos, singular and plural forms and compounded tags. Levenhstein distance (measures the number of required changes: substitution, insertion and deletion of a character are allowed operations to transform one tag into another) is computed for each tag pair from the tag space. Tag space is divided into clusters The most frequent tag from each cluster is used for further computations

10 Improving coverage and diversity of clouds with clustering We explore 3 different clustering techniques, each obtained cluster expresses a laten topic in the tag space. The goal is to cover as many clusters as possible. 1 A tag space is clustered and divided into disjoint clusters. 2 Tags are selected proportionally from each cluster according to their coverage. 3 The tags selection algorithm maximizes a chained coverage metric. The maximization of the chained coverage promotes (specific) tags with the high coverage of not yet covered documents by previously selected tags. On the other hand frequent tags with low chained coverage (general meaning) are omitted.

11 Experiments The improvements of the methodologies are evaluated in terms of coverage and overlap of generated tag clouds. The evaluation is conducted on the following datasets: Bibsonomy contains 5794 distinct users, items and tags. The total number of tagging posts is Delicious dataset contains users, unique tags and bookmarks. The total number of tagging posts is

12 Syntactical pre-clustering of tags Table: Mean values of coverage and overlap for the baseline and syntactical pre-clustering methods on BibSonomy and Delicious datasets. Coverage Overlap Dataset Baseline Pre-Clustering Baseline Pre-Clustering BibSonomy Delicious Coverage had a 5% (5079 documents) increase on BiSonomy dataset and 3.5% (3072 documents) increase on Delicious. Overlap, on the other hand, had similar results.

13 Syntactical pre-clustering of tags As number of tags increases, coverage and overlap improves in a logarithmic fashion. Baseline Syntactical clustering Coverage Overlap Delicious Number of tags in the tag cloud Delicious Number of tags in the tag cloud Coverage Overlap Bibsonomy Number of tags in the tag cloud Bibsonomy Number of tags in the tag cloud Figure: Coverage and overlap results for baseline (red) and pre-clustering (black) methods and their corresponding logarithmic fit.

14 Improving coverage and diversity of clouds with clustering The average improvements are presented in the following tables. Coverage Dataset Baseline K-means Hierarchical Feature hashing BibSonomy Delicious Table: Mean values of coverage for the baseline and different clustering methods on BibSonomy and Delicious datasets. Overlap Dataset Baseline K-means Hierarchical Feature hashing BibSonomy Delicious Table: Mean values of overlap for the baseline and different clustering methods on BibSonomy and Delicious datasets.

15 Improving coverage and diversity of clouds with clustering The proposed methodology improves the coverage on both datasets. Similarly, the overlap of generated tag clouds is decreased. The best performing clustering technique is hierarchical clustering which computes a tag pairs co-occurrences. Coverage Overlap Delicious Number of tags in the tag cloud Delicious 0.1 Baseline K-means Hierarchical Feature hashing Number of tags in the tag cloud Coverage Overlap Bibsonomy Number of tags in the tag cloud Bibsonomy Number of tags in the tag cloud Figure: Improvements of coverage and overlap on Bibsonomy and Delicious datasets with different clustering techniques and their corresponding logarithmic fit.

16 Conclusion and future work Syntactical pre-clustering of tags improves coverage of tag clouds it prohibits a depiction of the syntactically similar tags = more diverse tag clouds Second methodology improves coverage and decreases overlap of tag clouds introduced metric chained coverage simplifies a selection process As a future work we intend to explore possible new metrics that would incorporate well-known metrics altogether and in a such way simplify a selection process of tags.

M-Eco enhanced Adaptation Service (D5.2) Dolog, Peter; Durao, Frederico Araujo; Lage, Ricardo Gomes; Leginus, Martin; Pan, Rong

Aalborg Universitet M-Eco enhanced Adaptation Service (D5.2) Dolog, Peter; Durao, Frederico Araujo; Lage, Ricardo Gomes; Leginus, Martin; Pan, Rong Publication date: 2012 Document Version Accepted author