1 Data Mining over the Deep Web Tantan Liu, Gagan Agrawal {liut,agrawal}@cse.ohio-state.edu Ohio State University April 12, 2011.

1

Data Mining over the Deep Web

Tantan Liu, Gagan Agrawal {liut,agrawal}@cse.ohio-state.edu

Ohio State University

April 12, 2011

2

Outline

• Introduction– Deep Web– Data Mining on the deep web

• Frequent itemset mining over the deep web– Bayesian network– Active learning based sampling method

• Experiment Result• Conclusion

3

Deep Web

• Data sources hidden from the Internet– Online query interface vs. Database– Database accessible through online Interface– Input attribute vs. Output attribute

• An example of Deep Web

4

Data Mining over the Deep Web

• High level summary of data– Scenario 1: A student wants to find a job as a software En

gineer• Will a master degree help?

• Which language to learn: Java, C, or C#?

• Try MSN careers – to much information!

• Frequent itemset mining!

5

Challenges

• Databases cannot be accessed directly– Sampling method for Deep web mining

• Obtaining data is time consuming– Efficient sampling method

– High accuracy with low sampling cost

6

Roadmap

• Introduction– Deep Web– Data Mining

• Frequent Itemset mining over the deep web – Bayesian Network– Active learning based sampling method

• Experiment Result

• Conclusion

7

Frequent Itemset Mining

• Itemset: a set of attributes with instantiations, e.g I={Brand=benz, Age>5}• Support(Brand=Benz, Age>5)=2/8=0.25• Frequent Itemset: Support is larger than a threshold

8

Frequent Itemset Mining on Deep Web

• Challenges

– Support of itemsets is unavailable

– The size of itemsets could be huge• Considering 1-itemsets

– Simple random sample – Inefficient• Support of itemsets of input attributes is known

– # of data records satisfying the query is provided

9

Main Idea

• Task: Estimating the support of itemsets of output attributes• Questions

– Can we use information about input attributes?• Bayesian Network

– Relation between input attributes and output attributes

– Compute support for itemsets of output attributes

– How to quickly build the model• Active learning based sample method

10

Bayesian Network

• Relation between input and output attributes• Graphical model

– Random variables • Input and output attributes

– Conditional dependencies• Output attributes depend on input attributes

11

Active Learning

• In machine learning– Passive learning: data are randomly chosen – Active Learning

• Certain data are selected, to help build a better model

• Active Learning– Obtaining data is costly and/or time-consuming

• Frequent Itemset Mining on Deep Web

12

An Example of Bayesian Network

Brand Age

Mileage Price Price

<=5000 >5000 Brand Age

H <=5 0.5

0.5

0.0

0.0

H >5

B <=5

B >5

• Support of Itemsets depends on parameters in the Bayesian network

• Parameters are estimated based on Sample

‒ Parameter: p(price<=5000|H,<=5)

• 2 data records satisfying brand=H, Age<=5

• 1 data records satisfying brand=H, Age<=5, price<=5000

‒ p(price<=5000|H,<=5)=1/2=0.5

p

Support(Price<=5000)= 0.25

H, B <=5, >5

known Estimate!

[0.125 0.125 0.0 0.0]

0.25

0.25

0.25

0.25

0.5

0.5

1.0

1.0

0.25

0.25

0.25

0.25

0.5

0.5

0.0

0.0

13

Example of Active learning on Deep Web

Deep Web Data Source

B=H&Age<=5

B=H&Age>5

B=B&Age<=5

B=B&Age>5

Price

Q1 Q2 Q3 Q4

Q1Q2Q3Q4

Price

<=5000 >5000p11 p12 p21 p22p31 p32p41 p42

Qi, i=1,…, 4 Sampled Data

14

An Example of Active Learning Based Sampling

• Hidden idea– Sampling heavily on query spaces with high impurity

Q1(B=H) Q2(B=B)

Price 0.01 0.99 0.5 0.5

Q2

Q1

Deep Web Data Source

15

Detailed Formulation

• Support for output attributes

‒ : an instantiation of input attributes, or a query– : prior probability

• Known

– , • Conditional probability• Parameters in conditional table

• Unknown, need to estimate

16

Parameters in Bayesian Network

• are estimated based on a sample

• Difference between estimated values and true values

• Consider as statistical variables

• Conjugate distribution– After observing data D, is in the same family with

• Hyper parameter – Expectation: , where

• Estimation for support of output attributes– Expectation on the distribution

17

Active Learning on Deep Web

• Risk Function– Risk with the estimation for 1-itemsets composed of output attributes– Based on the hyper parameter in the Bayesian Network,

• Data Selection– Data are obtained by queries : query selection– Data records are selected step by step– Choosing the query with most reduction on risk function

• Updating Model– For , and sample

where denotes the number of data records containing

18

Support for n-itemsets(n>1)

• Estimation based on the Bayesian network

– • Support value of in the query space

19

Roadmap

• Introduction– Deep Web– Data Mining

• Frequent Itemset mining over the deep web – Bayesian Network– Active learning based sampling method

• Experiment Result

• Conclusion

20

Experiment Result

• Data set: US census– 2008 US Census on the income of US households– 40,000 data records

• Three Methods– Dir:

• Random Sample• Direct Computation

– Bay• Random Sample• Computation Based on Bayesian Network

– Act: our proposed method• Active Learning based Sample• Computation Based on Bayesian Network

21

US census

• Square Error Rate:

• Absolute Error Rate (AER):

22

Conclusion

• Data mining on the deep web is challenging• Frequent itemset mining over the deep web• Bayesian network is used to model the deep web• A active learning based sampling method• The experiment results show the efficiency of our work

23

Questions?