Semantic Query By Example

1
Semantic Query By Example

• Lipyeow Lim, Haixun Wang, Min Wang
• IBM T.J. Watson Research Center
• Jiang Kai jiangkai_at_fudan.edu.cn

2
Outline

• Introduction
• Background
• Semantic Query By Example
• Feature Extraction
• Learning Query Semantics
• Active Learning for QBE
• Experiments
• Conclusion

3
Introduction

• Urgent need to incorporate ontology into the
  realm of RDBMS to support semantic queries
• Difficulties in semantic query on ontology
• Integrate data and its related ontology
• Express semantic queries

4
Examples EMR

• Different disease codes for the same symptoms
• Iris Neoplasm, Tumor of the Uvea, Eye
  Neoplasm

5
(No Transcript)
6
Example

• Query find all patients diagnosed with eye tumor
• RDF-like data model, store ontology in
  Thesaurus(src, rel, tgt)

7
Example

• XML format

8
Challenge

• Hard to write the query and needs to have an
  intimate knowledge of the structure of the
  ontology
• Fuzziness in the semantics of the query
• Eg. Find all patients diagnosed with some disease
  in the choroid (???) , which is part of eye.
  Relationships linking disease concepts to
  anatomic locations Disease_Has_Primary_Anatomic_S
  ite, Disease_Has_Associated_Anatomic_Site,
  Disease_Has_Metastatic_Anatomic_Site

9
Contributions

• A novel approach to address the problem
• A novel QBE (Query By Example) method
• Active learning to improve the accuracy of query
  semantics modeling
• Experimentally validate the effectiveness of the
  method with as few as 2-4 user-given examples

10
Review on QBE

• First proposed in 1970s, two common
  characteristics
• The examples are used to specify a query that
  will be generated
• The generated query is a normal query in terms
  that all the query conditions are defined on the
  base attributes in the underlying tables in the
  database
• The novel semantic QBE framework
• Learn the query processing directly without
  learning the query first
• Not only on the base attributes but also
  semantics of the base data encoded in the
  ontology and the connections between them

11
Background

• Supervised Learning
• Create a function f from a training dataset
• Predict any valid input object x as yf (x)
• Accuracy depends largely on the quality of the
  training dataset
• Active Learning
• Situations in which unlabeled data is abundant
  but labeling data is expensive
• Query the user for labels of examples
• Graph Classification
• Classify object (graph/subgraph) based on a set
  of labeled graphs
• Classify vertices on a graph
• Vertices and edges in the ontology have their own
  content

12
Semantic Query by Example

• Query posed against a base table D(X,OID)
• Offline process
• Compute the feature vector associated with OID
• Online process
• User poses a query by giving a set of example
  tuples. SVM learn a model from the examples
• User determine whether predictions is accurate
  for an additional small no. of tuples

13
Feature Extraction

• Challenges
• Feature space must capture enough information so
  that all discriminative features are covered
• Vectorize the local hierarchical structure of an
  ontology node
• Naïve scheme include immediate child nodes
  parent nodes, paths to some important nodes
• Nodes and paths still complex
• Not general enough
• Alternative perform feature extraction per query
• Expensive
• Require user to know what part of ontology are
  important

14
Shortest distance based feature extraction

• Ontology has a set of concepts Cc1,ck.
  f(OID)lty1,ykgt, yi is the shortest distance from
  node o to any node whose concept is ci.
• Omits much information in the ontology
• Treat the ontology as an undirected graph
• Disregard the labels of the edges, concentrate
  only on the labels of the nodes
• Effectively capture the local structures
• Tell how close a node is to different concepts
• Infer much hierarchical information since the
  distribution of concepts is not uniformly random

15
Example
16
Example

• Find nodes that are parent nodes of any node of
  concept C

17
Implementation
18
Learning Query Semantics

• User provides examples following format (X,
  OID, label), label ?1, -1
• Randomly pick a set of tuples Pn in D and label
  them as negative.
• Tuples satisfying the query is a small
  proportion, classifier usually can overcome such
  noise

19
Support Vector Machine

• SVMs learn a linear decision boundary to
  discriminate between the two classes.
• SVM trained from dataset (xi, labeli) specifies a
  linear classification rule by a pair (w, b),
  where w ?RN, f(x) wx b, where x is
  classified as positive if f(x)gt0, or negative if
  f(x)lt0.
• Decision boundary x ?RN wx b0

20
SVM

• Optimal boundary

21
Active Learning for QBE

• Iteratively select the best example to
  incorporate into the training dataset to obtain a
  high quality classifier

22
Experiment

• Queries

23
Varying the no. of positive examples

• Table size1000, training dataset 40
• Query selectivity 10

24
Varying the num of training examples

• Num of positive examples 8
• Query selectivity 10

25
Varying the num of training examples
26
Varying query selectivity

• Num of training examples 160
• Num of positive examples 8

27
Varying table size

• Num of training examples 160
• Num of positive examples 16
• Query selectivity 5

28
Processing time

• Processing a query by example on a table of size
  10,000 takes only 0.25 sec

29
Conclusion

• Introduce a data mining approach to support
  semantic queries in relational database
• Dealing with the ontology directly when asking
  semantic queries
• Efficient, effective and general in supporting
  semantic queries

30
The End

• Thank you all!