A New Suffix Tree Similarity Measure for Document Clustering Hung Chim, Xiaotie Deng WWW 07.

A New Suffix TreeSimilarity Measure forDocument Clustering

Hung Chim, Xiaotie Deng

WWW 07

1. Document Clustering

• Agglomerative Hierarchical Clustering (AHC)

• Suffix Tree Clustering (STC)

- commonly used in result clustering

2-1. Suffix Tree Clustering

Ex: 3 documents

• cat ate cheese• cat ate mouse too• mouse ate cheese too

cat ate cheese

cat ate cheese

cat ate cheese

cat ate cheese

score(B) = |B| f(|P|)f: remove stopwords, <= 3

, > 40% && penalize single word, constant for |P| > 6

2-2. Base Cluster

2-3. Combining Base Cluster

• Keep top k(=500) base cluster

• Merge high overlap base clustersmerge Bi & Bj iff

|Bi∩Bj| / |Bi| > 0.5

|Bj∩Bi| / |Bj| > 0.5

2-4. Advantage

• High precision even using snippet

• Incremental and linear time

• Order Independent

• No magic k

top k base clusters? 0.5?

3. New Suffix Tree Clustering

diT =

[tfidf(n1, di), tfidf(n2, di), …]

Group-average AHC

(GAHC)

4. Evaluation

• Use F-measure

precision(Ci, Gj) = |Ci∩ Gj | / |Ci|

recall(Ci, Gj) = |Ci∩ Gj | / | Gj |

• OHSUMED Document CollectionMeSH indexing terms

• RCV1 Document Collectioncategories

5. Comparison

• STC : seldom generate large cluster

• NSTC : not incremental