SourceRank: Trust and Relevance based Ranking of
Web Databases for the Deep Web
Raju Balakrishnan, Subbarao Kambhampati
rajub@asu.edu
rao@asu.edu
Evaluated on Google base and online Sources.
Comparison with Google Product Search
Source Selection in Deep Web
0.78
Millions of
Databases: Which
are Trustworthy and
Relevant?
S3
Mediator
S1
0.86
0.3
0.25
Online Sources
Precision
DCG
0.2
0.22
Web DB
0.15
Web DB
0.4
0.14
Web DB
0.6
Web DB
Web DB
0.05
Deep Web
0
W (S1
S 2)
(1
Agreement Implies Trust & Relevance
SourceRank
(SR)
CORI
SR-CORI
SR-Coverage
Google Base Books Precision
0.5
0.4
0.3
t1
Godfather Paperback Used $9.99
0.3
0.7
0
0
.
.
9
1
0.4
0.7
Relevant
&True (RT)
t
n
a
v
)
e
l
R
e
(
R les
p
u
T
Coverage
S2
A( R1, R 2)
)
| R2 |
SourceRank is calculated as the stationary visit
probability of a weighted random walk on the database
vertex in the agreement graph.
Universal Set
of Tuples (U)
0.1
Tuple Model
and Record
Linkage
0.2
0.1
t2
0
R1
R2
)
T
e
(
u les
r
T p
u
T
R3
Let t1, t 2 RT, f 1, f 2 U RT and P a be the agreement
probability. Assuming independent selection,
U
Paperback The Godfather $13.99
RT
Pa (t1, t 2)
Pa ( f 1, f 2)
Agreement between the sources is modeled as
an Agreement Graph.
Computing agreement requires record linkage
Value similarity : SoftTFIDF with Jaro-Winkler.
Attribute importance is calculated based on mean
inverse document frequency of the token values.
Domain Independent.
Predefined schema mapping is not assumed.
 Quadratic Time complexity on Sources.
Sampling Databases
 Graph is computed based on samples from the
databases.
 Non-Cooperative Query Based Sampling.
URL: http://factal.eas.asu.edu
GBase
GBase-Domain
SourceRank
Coverage
Trustworthiness is evaluated as the decrease in ranks of
corrupted sources.