“Artificial Intelligence” in my research

Seung-won HwangDepartment of CSE

POSTECH

2

Recap

Bridging the gap between under-/over-specified user queries

We went through various techniques to support intelligent querying, implicitly/automatically from data, prior users, specific user, and domain knowledge

My research shares the same goal, with some AI techniques applied (e.g., search, machine learning)

3

The Context:

Rank Formulation

Rank Processing

select * from housesorder by [ranking function F]limit 3

ranked results

querytop-3 houses

e.g., realtor.com

4

Overview

Rank Formulation

Rank Processing

select * from housesorder by [ranking function F]limit 3

ranked results

querytop-3 houses

e.g., realtor.com

Usability:Rank Formulation

Efficiency:Processing Algorithms

5

Part I: Rank Processing

Essentially a search problem (you studied in AI)

6

Limitation of Naïve approach

a:0.90, b:0.80, c:0.70, d:0.60, e:0.50

b:0.78 Algorithm

F = min(new,cheap,large)k = 1

Sort stepMerge stepnew (search predicate) : x

cheap (expensive predicate) : pc

large (expensive predicate) : pl

d:0.90, a:0.85, b:0.78, c:0.75, e:0.70

b:0.90, d:0.90, e:0.80, a:0.75, c:0.20

Our goal is to schedule the order of probes to minimize the number of probes

7

a:0.9

b:0.8

c:0.7

d:0.6

e:0.5

a:0.85

b:0.8

c:0.7

d:0.6

e:0.5

pr(a,pc) =0.85

pr(a,pl) =0.75

OID x pc pl min(x, pc, pl)

a 0.90

b 0.80

c 0.70

d 0.60

e 0.50

0.85

0.78

0.75

0.90

0.75

0.78

global schedule : H(pc, pl)

Unnecessary probes

initial state

a b c d

ea b c d

e

b

bgoal state

8

Search Strategies?

Depth-first Breadth-first Depth-limited / iterative deepening (try every depth

limit) Bidirectional Iterative improvement (greedy/hill climbing)

9

Best First Search

Determining which node to explore next, using evaluation function

Evaluation function: exploring more on object with the highest “upper bound

score” We could show that this evaluation function minimizes the

number of evaluation, by evaluating only when “absolutely necessary”.

10

Necessary Probes?

Necessary probes probe pr(u,p) is necessary if we cannot determine top-k

answers until probing pr(u,p), where u: object, p: predicate

OID x pc pl min(x, pc, pl)

a 0.90

b 0.80

c 0.70

d 0.60

e 0.50

top-1: b(0.78)

Can we decide top-1 without probing pr(a,pc)?

0.85

0.78

0.75

0.75

0.90

0.20

0.75

0.78

0.20

≤0.90

Let global schedule be H(pc, pl)

No pr(a,pc) necessary!

0.90 0.90

0.70 0.80

0.60

0.50

11

a:0.9

b:0.8

c:0.7

d:0.6

e:0.5

a:0.85

b:0.8

c:0.7

d:0.6

e:0.5

b:0.8

a:0.75

c:0.7

d:0.6

e:0.5

a:0.75

c:0.7

d:0.6

e:0.5

b:0.78b:0.78

a:0.75

c:0.7

d:0.6

e:0.5

b:0.78

pr(a,pc) =0.85

pr(a,pl) =0.75

pr(b,pc) =0.78

pr(b,pl) =0.90

Top-1

OID x pc pl min(x, pc, pl)

a 0.90

b 0.80

c 0.70

d 0.60

e 0.50

0.85

0.78

0.75

0.90

0.75

0.78

global schedule : H(pc, pl)

Unnecessary probes

12

Generalization

FA, TA, QuickCombine

r =1 (cheap)

r = h (expensive)

r = (impossible)

CA, SR-Combine

NRA, StreamCombine

s =1 (cheap)

s = h (expensive)

s = (impossible)

Random AccessSorted Access

FA, TA, QuickCombine

NRA, StreamCombine

MPro[SIGMOD02/TODS]

Unified Top-k Optimization[ICDE05a/TKDE]

13

Strong nuclear force

Electromagnetic force

Weak nuclear force

Gravitational force

Unified field theory

Just for Laugh: Adapted from Hyountaek Yong’s presentation

14

FA

TA

NRA

CA

MPro

Unified Cost-based Approach

15

Generality

Across a wide range of scenarios One algorithm for all

16

Adaptivity

Optimal at specific runtime scenario

17

Cost based Approach

Cost-based optimization Finding optimal algorithm for the given scenario, with

minimum cost,

from a space

)(argmin Ω MCostM Mopt

Mopt

18

Evaluation: Unification and Contrast (v. TA)

T

NN

T

N

Unification: For symmetric function, e.g., avg(p1, p2), framework NC behaves similarly to TA

Contrast: For asymmetric function, e.g., min(p1, p2), NC adapts with different behaviors and outperforms TA

depth into p1

dept

h in

to p

2

depth into p1

dept

h in

to p

2

costcost

19

Part II: Rank Formulation

Rank Formulation

Rank Processing

select * from housesorder by [ranking function F]limit 3

ranked results

querytop-3 houses

e.g., realtor.com

Usability:Rank Formulation

Efficiency:Processing Algorithms

20

Learning F from implicit user interactions

Using machine learning technique (that you will learn soon!) to combine quantitative model for efficiency and qualitative model for usability

Quantitative model Query condition is represented as a mapping F of objects into

absolute numerical scores DB-friendly, by attaining the absolute score on each object Example F( )=0.9 F( )=0.5

Qualitative model Query condition is represented as a relative ordering of

objects User-friendly by alleviating user from specifying the absolute

score on each object Example >

21

A Solution: RankFP (RANK Formulation and Processing)For usability, a qualitative formulation front-end which enables rank

formulation by ordering samples

For efficiency, a quantitative ranking function F which can be efficiently processed

sample S (unordered)

Sample Selection:generate new S

Function Learning:learn new F

ranking R* over S Over S:RF R* ?

12345

no

yes

F

ranking function

Rank Formulation Rank Processing

ranked results

processing of Q

Q: select * from houses order by F limit k

22

Task 1: RankingClassification

Challenge: Unlike a conventional learning problem of classifying objects into groups, we learn a desired ordering of all objects

Solution: We transform ranking into a classification on pairwise comparisons [Herbrich00]

learning algorithms: a binary classifier

+

-F

a-b

b-cc-dd-ea-c… …

ranking view: c > b > d > e > a

cb

de

a

classification view:

--++-

pairwise comparison classification

[Herbrich00] R. Herbrich, et. al. Large margin rank boundary for ordinal regression. MIT Press, 2000.

23

Task 2: ClassificationRanking

Challenge: With the pairwise classification function, we need to efficiently process ranking.

Solution: developing duality connecting F also as a global per-object ranking function.

Suppose function F is linearClassification View: Ranking View:F(ui-uj)>0 F(ui)- F(uj)>0 F(ui)> F(uj)

b

d e

ac

F(a-b)? F(a)=0.7

F(a-c)?

F(a-d)?….. Rank with F(.)

e.g., F(c)>F(b)>F(d)>…

24

Task 3: Active Learning

Finding samples maximizing learning effectiveness Selective sampling: resolving the ambiguity

Top sampling: focusing on top results

Achieving >90% accuracy in <=3 iterations (<=10 ms)

F

F

25

Using Categorization for Intelligent Retrieval

Category structure created a-priori (typically a manual process) At search time: each search result placed under pre-assigned category Susceptible to skew information overload

26

Categorization: Cost-based Optimization

Categorize results automatically/dynamically Generate labeled, hierarchical category structure dynamically

based on the contents of the tuples in the result set

Does not suffer from problems as in a-priori categorization

Contributions: Exploration/cost models to quantify information overload

faced by an user during an exploration

Cost-driven search to find low cost categorizations

Experiments to evaluate models/algorithms

27

Thank You!