QPIAD: Query Processing over Incomplete Autonomous Databases

1
QPIAD Query Processing over Incomplete
Autonomous Databases

• Hemal Khatri (Arizona State University)
• Jianchun Fan (Arizona State University)
• Garrett Wolf (Arizona State University)
• Yi Chen (Arizona State University)
• Subbarao Kambhampati (Arizona State University)

2
Incompleteness in Web databases

• Population by Lay Users

Automated Extraction
3
Problem
How to retrieve ranked relevant uncertain answers
for user queries?

• Challenges
• How to retrieve relevant uncertain answers
  through form-based interfaces of autonomous
  databases?
• How to keep query processing cost manageable?
• How to rank the retrieved uncertain answers?

• Possible Approaches
• QBody style Convt
• 1.CERTAIN ANSWERS ONLY Return certain answers
  only as in traditional databases (Low Recall)
• 2. ALL RETURNED Return certain answers and
  answers having body style value missing (Low
  precision, Infeasible)
• 3. ALL RANKED Ranking all answers by predicting
  values of missing attribute (Costly, Infeasible)

4
QPIAD System Architecture
5
Retrieving Relevant Answers via Query Rewriting
Given a query Q(Body styleConvt) retrieve all
relevant tuples
Base Result Set
Q
AFD Model Body style
Use F-measure to select top K Rewritten
Queries Q1 ModelA4 Q2 ModelZ4 Q3
ModelBoxster
Re-order top K queries based on Estimated
Precision
Ranked Relevant Uncertain Answers
F-Measure (1a)PR/(aPR) P Estimated
Precision R Estimated Recall based on P and
Estimated Selectivity
6
Learning Statistics to support Ranking Rewriting

• Learning attribute correlations by Approximate
  Functional Dependency(AFD) and Approximate
  Key(AKey)

Determining Set(Y) dtrSet(Y)
Sample Database
Prune based on AKEY
TANE
AFDs (XY) confidence

• Learning value distributions using Naïve Bayes
  Classifiers(NBC)

Learn NBC classifiers with m-estimates
Determining Set(Am)
Feature Selection
Estimated Precision P(AmvmdtrSet(Am))

• Learning Selectivity Estimates of Rewritten
  Queries(QSel) based on
• Selectivity of rewritten query issued on sample
• Ratio of original database size over sample
• Percentage of incomplete tuples while creating
  sample

7
Empirical Evaluation
Two experimental databases Cars(Cars.com) and
Census(UCI ML)

• Experimental Setup
• Oracular study used to measure Precision/Recall
  by artificially introducing missing values in
  databases.
• AFDs and NBC classifiers learned for various
  samples sizes ranging from 3 to 15.
• Purpose of Experiments
• Measuring quality of uncertain results returned
  by QPIAD Figure 1.
• Efficiency of QPIAD in retrieving relevant
  results Figure 2.
• Robustness of the learning algorithms used in
  QPIAD wrt to various sample sizes Figure 3.

Figure 1
Figure 3
Figure 2
8
QPIAD Web Interface