Intelligent Database Systems Lab N.Y.U.S.T. I. M. Supporting personalized ranking over categorical attributes Presenter : Lin, Shu-Han Authors : Gae-won.

Intelligent Database Systems Lab

N.Y.U.S.T.I. M.

Supporting personalized ranking over categorical attributes

Presenter : Lin, Shu-Han

Authors : Gae-won You, Seung-won Hwang, Hwanjo Yu

Information Sciences 178(2008)


N.Y.U.S.T.I. M.

2

Outline

Motivation Objective Methodology Experiments Conclusion Comments


N.Y.U.S.T.I. M.Motivation

3

Categorical attributes’ problem of information retrieval's personal ranking

Categorical attributes do not have an inherent ordering. How to rank the relevant data by categorical attribute.

For example, how can we…

Find old female with the preference of soda drink.

Name age Gender FavoriteDrink

Buy

Jane 30 Female Coke Coke, Milk

Mary 25 Female Pepsi Coke, Pepsi

Tom 21 Male Water Milk, Water

Denny 26 Male Coke Milk, Juice

Tina 11 Female Pepsi Red Wine, Pepsi


N.Y.U.S.T.I. M.Objectives

Enable a uniform ranked retrieval over a combination of categorical attributes and numerical attributes.

Support ranking of binary representation of categorical attribute Binary encoding

Sparsity

4

Name Female

Jane 1

Mary 1

Tom 0

Denny 0

Tina 1

Name Coke Pepsi Water

Jane 1 0 0

Mary 0 1 0

Tom 0 0 1

Name Coke Pepsi Water

Jane 1 0 1

Mary 1 1 0

Tom 1 0 1

Multi-valued attribute with bounded cardinality (item set, bc=2)Single-valued attribute


N.Y.U.S.T.I. M.Overview

5

（ 1）

（ 2）

（ 3）


N.Y.U.S.T.I. M.Rank formulation

6F= 0.5*age + 3*female + …


N.Y.U.S.T.I. M.Rank processing (TA)

A Simple example query:Find old female with the preference of soda drink.

Transform into

F= age + female

1. Candidate identification1. Sorted Access age and female

2. Find top-k sa(age) and sa(female), e.g., k=1, sa(age)={o1}; sa(female)={o2}

2. Candidate reduction1. O1=30+0

2. O2=25+1

3. O1 with the highest F score

3. Termination1. O1 !> F(30,1)=31 // upper bound score

2. Another round of sorted access to consider more candidates, e.g., sa(age)={O4}; sa(female)={O3}

7


N.Y.U.S.T.I. M.Bitmap – binary encoding

F=v1+v2+v3+v4, k=2

1) K={}, C={1111}（ Initailization）2) OID=excute(C)

3) OID={o4},|OID|>0,K={[o4,4]}

4) C={0111/1011/1101/1110} （ Expansion）5) K.count < k, Back to 2)

6) …

8

v1 v2 v3 v4

O1 1 0 1 1

O2 0 1 0 0

O3 0 1 1 1

O4 1 1 1 1

o5 1 0 1 1


N.Y.U.S.T.I. M.Bitmap– sparsity

Single-valued attributeF=w1v1+w2v2+…+w6v6

ranked weightw1 w2 w3;w4 w5 w6≧ ≧ ≧ ≧for simple, all w=1,k=2

1) K={}, C={100.100.100} （ Initailization）2) OID=excute(C)

3) OID={o4},|OID|>0, K=OID={[o4,2]}

4) C={010.100.100/ 100.010.100/100.100.010} （ Expansion）5) K.count<k, Back to 2)

6) …

9

　Attribute1 Attribute2 Attribute2

v1 V2 V3 V4 V5 V6 V4 V5 V6

O1 1 0 0 0 0 1 0 1 0

O2 0 1 0 0 1 0 1 0 0

O3 0 1 0 1 0 0 0 1 0

O4 1 0 0 1 0 0 1 0 0

o5 0 0 1 0 0 1 0 1 0


N.Y.U.S.T.I. M.Bitmap– sparsity

Multi-valued attribute with bounded cardinality

10

Attribute1 Attribute2

v1-1 V1-2 V1-3 V1-4 V2-1 V2-2 V2-3

O1 1 0 0 1 1 0 1

O2 0 1 0 1 1 1 0

O3 0 1 1 0 1 1 0

O4 1 0 0 1 1 0 1

o5 0 0 1 1 0 1 1


N.Y.U.S.T.I. M.Experiments

11

• UCI’s sparsity of indicating variable

• 22% of dataset consist only the categorical attributes.

• 56% of combination of numerical & categorical attributes.


N.Y.U.S.T.I. M.Experiments – synthetic data

12


N.Y.U.S.T.I. M.Experiments – real-life data

13


N.Y.U.S.T.I. M.

14

Conclusions

This paper studies How to support rank formulation

Processing over data with categorical attributes

Instead of adopting existing numerical algorithms, develop a bitmap-based approach to Binary encoding Sparsity

Single-valued Multi-valued with bounded cardinality


N.Y.U.S.T.I. M.

15

Comments

Advantage …

Drawback …

Application …

Intelligent Database Systems Lab N.Y.U.S.T. I. M. Supporting personalized ranking over categorical attributes Presenter : Lin, Shu-Han Authors : Gae-won.

Documents