A Comprehensive Comparison Study of Document Clustering for a Biomedical Digital Library MEDLINE

1. A Comprehensive Comparison Study of Document Clustering for a Biomedical Digital Library MEDLINE Advisor : Dr. Hsu Graduate : Jing Wei Lin Authors :lllhoi Yoo and Xiaohua Hu 2006.JCDL.10

2. Outline • Motivation • Objective • Use MeSH Ontology on Vector Space Model • Classification of Document Clustering Approaches • Experiment Results • Conclusion

3. Motivation • Document clustering has been used for better document retrieval, document browsing, and text mining in digital library.

4. Objective • The goal of this paper is to perform a comprehensive comparison study of various document clustering approaches in terms of the efficiency, the effectiveness, and the scalability.

5. Use MeSH Ontology on Vector Space Model • Problem : • Each document d is represented as a high dimensional vector of words/terms frequencies. • Solution :For example, “Neoplasms” as a descriptor has the following entry terms {“Cancer”, “Cancers”, “Neoplasm”, “Tumors”}.

6. Classification of Document Clustering Approaches • Hierarchical clustering algorithms • Advantage: they generate a document hierarchy so that users can drill up and drill down for specific topics of interest. • Disadvantage: because of their cubic time complexity, they are very much limited for very large documents. • Partitional clustering algorithms • Advantage: their time complexity is lower, for example, K-means is O(k*T*n). • Disadvantage: clustering results are heavily sensitive to the initial centroids because the centroids are randomly selected.

7. CL CL w w CR CR …. …. CL w CR

8. Experiment Results • How much does the selection method of Bisecting K-means for clusters to be bisected affect clustering results?

9. Experiment Results • Does Bisecting K-means outperform K-means?

10. Experiment Results • Does STC outperform hierarchical orpartitionalapproaches? STC v.s. Partitional STC v.s. Hierarchical

11. Experiment Results • Do partitional clustering algorithmsoutperform hierarchical clusteringalgorithms?

12. Experiment Results • Which of clustering algorithms is the most scalable or the least scalable?

13. Experiment Results • How much does a domain ontology MeSH improve document clustering?

14. Experiment Results

15. Experiment Results • How the clustering evaluation metrics are related to one another? • The smaller MI and Entropy, the better clustering quality while the bigger F-measure and purity, the better clustering quality.

16. Conclusion • The cluster selection methods of Bisecting K-means sufficiently affect clustering quality. • Bisecting Kmeans generally outperforms K-means if the cluster selection method Type A of Bisecting K-means is used. • STC provides better clustering solutions thanhierarchical algorithms but is worse than partitional clustering approaches.

17. Conclusion • Partitional clustering approaches are significantly superior to hierarchical approaches in terms of clustering evaluation metrics and the running times. • Bisecting K-means is normally superior to other clustering methods. • A domain ontology MeSH improves document clustering for MEDLINE articles.