Clustering XML Documents for Query Performance Enhancement

1
Clustering XML Documentsfor Query Performance
Enhancement

• Wang Lian

2
Outline

• Related Work
• Motivation
• Our Approach
• S-Graph
• Distance Function
• Clustering Algorithm
• Experimental Results

3
Related Work

• Besides storing XML documents in their native
  format, using RDBMS is an established trend.
• There are mainly two approaches for storing XML
  documents in RDBMS
• Schema-mapping
• Structure-mapping

4
Related Work(cont)

• In Schema-mapping
• A database schema is derived from the DTD of XML
  documents, therefore different DTD will generate
  different database schema
• In Structure mapping
• The database schema is fixed by defining a set of
  generic tables.

5
Motivation

• Using both schema or structure mapping, documents
  must be cut into pieces and inserted into tables.
  To answer a query tables should be joined to
  provide the answers.
• As the size of tables grows larger, the join cost
  may be very high.
• Our observation is if a collection contains
  documents of different structures, then
  clustering on documents structures may reduce
  the join cost.

6
Motivation(cont)

• An example
• DTD is

7
Motivation(cont)

• Documents in 3 clusters

8
Motivation(cont)

• Unpartitioned schema

9
Motivation(cont)

• Partitioned schema

10
Motivation(cont)

• Suppose we want to answer an Xpath query
  /conference/author.text()
• Using unpartitioned schema, table conference (2
  tuples )and author (9 tuples) should be joined,
• Using partitioned schema, table conference1 (2
  tuples ) and author1 (3 tuples) should be joined.

11
Our Approach

• XML document is a mixture of structure
  information and data value. In our context, only
  structure information is used to do clustering.
• We need a proper distance function before using
  any clustering algorithm.

12
S-Graph

• Given a set of XML documents C, the structure
  graph (s-graph) of C, sg(C) (N, E), is a
  directed graph such that N is the set of all the
  elements and attributes in the documents in C and
  (a, b)?E if and only if a is a parent element of
  b in document(s) in C (b can be element or
  attribute).
• Certainly, s-graph does not catch all structure
  information of documents, however it captures the
  parent-child relationship which is valuable for
  evaluating path expression.

13
Distance Function

• For two sets, C1 and C2, of XML documents, the
  distance between them,
• where sg(Ci) is the number of edges in
  sg(Ci), i1,2, and sg(C1)? sg(C2) is the set of
  common edges of sg(C1) and sg(C2).

14
Distance Function(cont)
Dist(doc1,doc2)1 and Dist(doc2,doc3)0.25
Tree-dist(doc1,doc2) Tree-dist(doc2,doc3
)1
15
Clustering Algorithm

• Input X the set of XML documents
• Input k the number of clusters specified by user
• SGpre-clustering(X)
• While(remaining cluster numbergtk)
• Merge cluster Ci and Cj which maximize a
  predefined goodness function

16
Clustering Algorithm(cont)

• Complexity
• n X, mSG,
• Time complexity
• The upper bound of pre-clustering is O(nm), in
  general, it can be reduced to O(n).
• Iterative merging O(m2logm)
• Space complexity
• O(m2)

17
Experimental Results

• The clustering algorithm is tested on real data,
  the DBLP XML records, which contains about
  200,000 documents composed by 36 elements.
• Pre-clustering is effective, after scan 200,000
  documents, only 233 distinguished s-graphes
  remain, which makes following clustering using
  only less than 2 second.

18
Experimental Results(cont)

• After setting the number of cluster to be 3, we
  get three clusters containing about 193,000
  documents, one for article, and the other two for
  inproceedings.
• The interesting thing is in those two clusters of
  inproceedings, one s-graph is a subgraph of
  another.

19
Experimental Results(cont)

• We use Oracle 8.1.5 to store all the documents in
  4 versions
• Version 1 unpartitioned schema-mapping
• Version 2 partitioned schema-mapping
• Version 3 unpartitioned structure mapping
• Version 4 partitioned structure mapping

20
Experimental Results

• Query type
• Q1 /A1/A2//Ak all possible absolute XPathes
  in the documents.
• Q2 /A1/A2//Aktext()value''/text()
• absolute XPaths in which Ai, i1,,k, are
  randomly
• picked and "value" is the value of Ak in some
  documents.
• Q3 /A1/A2//Akcontains(.,substring'')/text()
  
• same as Q2 except that the condition tested is
  Ak contains a substring

21
Experimental Results
22
Question?