Computer Science Technical Report Archive

USC Computer Science Technical Reports, no. 594 (1994)

(USC DC Other)

USC Computer Science Technical Reports, no. 594 (1994)

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content V o cabulary Problem in In ternet Resource Disco v ery
T ec hnical Rep ort USCCS
ShihHao Li and P eter B Danzig
Computer Science Departmen t
Univ ersit y of Southern California
Los Angeles California fshli danzig gcsuscedu
Abstract
When searc hing information in a retriev al system p eople use a v ariet y of terms to describ e
their information needs When the terms used in a query are dieren t from those indexed b y the
system users fail to obtain the information they w an t This is called the vo c abulary pr oblem This problem has b een studied and discussed in information retriev al for decades Recen tly
Deerw ester et al prop osed a new tec hnique based on singular v alue decomp osition and obtained
promising results In this pap er w e describ e ho w to apply this tec hnique to In ternet resource
disco v ery Index T erms information retriev al p olysem y resource disco v ery synon ym yv o cabulary prob
lem
In tro duction
When searc hing information in a retriev al system p eople use a v ariet y of terms to describ e their
information needs When the terms used in a query are dieren t from those indexed b y the system
users fail to obtain the information they w an t This is called the vo c abulary pr oblem In general there are t wot yp es of v o cabulary problem One is called the synonymy pr oblem the other the p olysemy pr oblem Both problems o ccur b ecause p eoples bac kgrounds and problem
con texts dier Synon ym y refers to the fact that p eople describ e the same information using
dieren t terms P olysem y refers to the fact that p eople use the same term for dieren t meanings
Tosolv e the v o cabulary problem Deerw ester et al prop ose L atent Semantic Indexing LSI
They assume there is some underlying seman tic structure in the pattern of term usage across
do cumen ts and use Singular V alue De c omp osition SVD to capture this structure LSI allo ws users
to searc h do cumen ts based on their concepts or meaning rather than simple query terms LSI has
b een tested on sev eral information systems with promising results LSI p erforms w ell in traditional information retriev al systems Ho w ev er it can not b e
directly applied to the In ternet where p eople searc h information from serv ers all o v er the w orld In
this pap er w e prop ose a new solution based on LSI to address the v o cabulary problem in In ternet
resource disco v ery Our metho d clusters serv ers according to their con ten ts and allo ws users to
select in terested serv ers using their fa v orite taxonomies

Laten t Seman tic Indexing
LSI is an extension of Saltons V e ctor Sp aceMo del VSM in whic h do cumen ts and queries are
representedasv ectors of the form
d
i
a
i a
i a
im
q
j
b
j b
j b
jm
where m is the n um b er of distinct terms in the database the co ecien ts a
iw
and b
jw
represen t
the w eigh t or frequency of term t
w
w m in do cumen t d
i
and query q
j
resp ectiv elyF or
a database of n do cumen ts it is represen ted as an m n termdo cumen t matrix The similarit y
bet w een d
i
and q
j
is based on the n um b er of their matc hing terms and is calculated b y the c osine
c o ecient Sim d
i
q
j
P
m
w a
iw
b
jw
q
P
m
w a
iw
P
m
w
b
jw
T o capture the seman tic structure among do cumen ts LSI applies SVD to an m n term
do cumen t matrix and generates v ectors of k t ypically to orthogonal indexing dimensions
where eac h dimension represen ts a linearly indep enden t concept The decomp osed v ectors are used
to representbothdocumen ts and terms in the same seman tic space while their v alues indicate
the degrees of asso ciation with the k underlying concepts Figure sho ws SVD applied to a term
do cumen t matrix
term-doc
matrix
db
term
matrix
(m xn)
(m xk)
document
matrix
(n xk)
SVD
(k)
Figure SVD applies to an m n termdo cumen t matrix where m and
n are the n um b ers of terms and do cumen ts in the database and k is the
n um b er of orthogonal indexing dimensions used b y SVD
A query v ector in LSI is the w eigh ted sum of its comp onentterm v ectors T o determine
relev an t do cumen ts the query is compared with all do cumen ts and those with the highest cosine
using cosine co ecien t are returned Because k is c hosen m uc h smaller than the n um ber of
terms and do cumen ts in the database ie the n um ber of ro ws and columns in the termdo cumen t
matrix those k concepts are neither term nor do cumen t frequencies but compressed forms of b oth
Therefore a query can hit do cumen ts without ha ving common terms but with common concepts
Example describ es the query pro cessing in b oth VSM and LSI

Example Let q d
i
i and t
j
j b e a user query a set of do cumen ts and their
asso ciated terms in an information system where
d
ft
t
t
g d
ft
t
t
g d
ft
t
t
t
g d
ft
t
g d
ft
t
t
g q ft
t
g T able sho ws their v ector represen tations and cosine similarities in VSM and LSI Figure sho ws
the dimensional plot of the decomp osed do cumen t and term v ectors in LSI
do c term VSM LSI
query description t
t
t
t
t
t
t
t
Sim dim dim Sim
d
t
t
t
d
t
t
t
d
t
t
t
t
d
t
t
d
t
t
t
q t
t
T able The v ector represen tations and similarities of do cumen ts d
i
i and query q in VSM and LSI The t
t
and dim dim
columns sho w the v ectors in VSM and LSI resp ectiv ely The Sim columns
giv e the cosine similaritybet w een eac h d
i
and q In traditional information retriev al systems a do cumen tis relev an t to a query if it con tains
all the terms in the queryF or example do cumen t d
is relev anttoquery q b ecause it con tains
b oth t
and t
in q In VSM a do cumen tis relev an t to a query if their similarit y is larger than
a predened threshold If the threshold is then d
is relev antto q b ecause Sim d
q Similar to VSM LSI uses cosine similarit y measure to determine the relev an t do cumen ts Using
the same threshold do cumen t d
d
and d
are relev an t to query q because Sim d
i
q for
i The ab o v e example sho ws that giv en the same threshold LSI returns more relev antdocu men ts than VSM This is b ecause the dimensions in decomp osed v ectors are not indep endentof
eac h other Therefore t wov ectors can b e relev an t without ha ving common terms
In ternet Resource Disco v ery
In a distributed en vironmen t p eople searc h information b y sending requests to asso ciated informa
tion serv ers using a clientserver mo del Figure a In this approac h the clien t needs to kno w

-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
2-D Plot of Terms and Documents
Dimension 1
Dimension 2
d1
d2
d3
d4
d5
t1
t2
t3
t4
t5
t6
t7
t8
q
Figure The dimensional plot of decomp osed v ectors in Example where
do cumen ts d
i
i terms t
j
j and query q are represen ted
as and resp ectiv ely The dashed cone represen ts the region
where do cumen ts suc has d
d
and d
are within a cosine of from
query q ie relev antto q a b
client directory
server server server
query
server(s)
result
request
client server
request
result
Figure a Clien tServ er Mo del b Clien tDirectoryServ er Mo del

the serv ers name or address b efore sending a request In the In ternet where thousands of serv ers
pro vide information it b ecomes dicult and inecien tto man ually searc h all the serv ers
One step to solv e this problem is to k eep a directory of services that records the description
or summary of eac h information serv er A user sends his query to the directory of services whic h
determines and ranks the information serv ers relev an t to the users request The user emplo ys the
rankings when selecting the information serv ers to query directlyW e call this the clientdir e ctory
server mo del Figure b
In ternet resource disco v ery services suc h as Arc hie W AIS WWW Gopher
and Indie allo w users to retriev e information throughout the In ternet All of the ab o v e
systems pro vide services similar to the clien tdirectoryserv er mo del F or example Arc hie W AIS
and Indie supp ort a global index lik e the directory of services in their systems WWW and Gopher
do not ha v e a global index b y themselv es Instead they ha v e an addedon indexing sc heme built
b y other to ols suc h as Harv est WWWW for WWW and V eronica for Gopher
The indexing sc hemes implemen ted in the ab o v e systems iden tify relev antserv ers bymatc h
ing k eyw ords b et w een their database represen tativ es and user queries They do not deal with the
v o cabulary problem This pap er describ es ho w to apply LSI in those systems
LSI in Clien tDi rectoryServ er Mo del
In a distributed en vironmen t eac h database has a dieren t set of do cumen ts and term usage Th us
the correlations represen ted b y the decomp osed v ectors in eac h database ha v e completely dieren t
meanings whic h results in a p oten tial problem of using LSI in In ternet resource disco v ery systems
If w e collect the do cumen t and term v ectors from all the databases at the directory of services
there is no w a y to determine the relev antserv ers for a user queryT osolv e this problem w e prop ose
the Se c ondL evel LSI and describ e ho w it handles the synon ym y and p olysem y problems
Synon ym y
When applying LSI in the clien tdirectoryserv er en vironmen t w e can collect the do cumentv ectors
from all the databases and p erform the SVD at the directory of services This approac h is easily
implemen ted but it suers from high comm unication o v erhead for transmitting all the do cumen t
v ectors from serv ers to the directory of services In addition the directory of services needs a h uge
space to store the do cumentv ectors from all the databases whic h is equiv alen t to the summation
of their space requiremen ts
T o alleviate this problem w e can summarize the do cumen ts of a database in a uniform w a y
that can b e easily understo o d and v eried b ymac hine Using existing clustering algorithms a
database can b e divided in to clusters where eac h cluster con tains do cumen ts ha ving a n um ber of
common terms The a v erage of all the do cumentv ectors in a cluster is used as its represen tativ e
also called cluster v ector The whole database is represen ted b y the union of all its cluster
represen tativ es in the form of a termcluster matrix Figure sho ws the termdo cumen t matrix
and its asso ciated termcluster matrix in Example The directory of services collects the cluster represen tativ es from all the databases and builds
a sup erset termcluster matrix These cluster represen tativ es are treated as individual do cumen ts
and decomp osed b y SVD The decomp osed term and cluster v ectors are used for query pro cess
ing just lik e the decomp osed term and do cumentv ectors in the original LSI Figure sho ws the

a b
d
d
d
d
d
t
t
t
t
t
t
t
t
c
c
c
t
t
t
t
t
t
t
t
Figure a The termdo cumen t matrix in Example and b its asso ci
ated termcluster matrix Clusters c
c
and c
consist of do cumen t sets
fd
d
d
g fd
gand fd
g resp ectiv ely The cluster represen tativeisthe
a v erage of its comp onen t do cumentv ectors
arc hitecture of SecondLev el LSI
P olysem y
In Deerw esters exp erimen ts LSI deals nicely with the synon ym y problem but only oers a
partial solution to the p olysem y problem The deciency results from the fact that a w ord with
m ultiple meanings is represen ted as a w eigh ted a v erage of all dieren t meanings where eac h meaning
is represen ted as a p oin t in the seman tic space
T o reduce the p olysem y problem in the clien tdirectoryserv er en vironmen t w e add m ultiple
taxonomies in to our system to help classifying query and do cumen t terms A taxonomyisa
predened classication sc heme suc h as the A CM classication system and the IEEE INSPEC
thesaurus F or eac h taxonomyw e generate a termpseudodo cumen t matrix where eac h
pseudo do cumen t is a set of subcategory names or synon ymous terms within the same category F or example in the A CM classication system
C Net w ork Op erations
Network management
Network monitoring
Public networks
C Distributed Systems
Distribute d applic ations
Distribute d datab ases
Network op er ating systems

term-cluster
matrix
db1
(m1 xn1)
term-cluster
matrix db2
(m2 xn2)
term-cluster
matrix db3
(m3 xn3)
term
matrix
(m xk)
cluster
matrix
(n xk)
SVD
(k)
term-cluster
matrix
(m xn)
Directory of Services
Server1
Server2
Server3
Figure SecondLev el LSI system arc hitecture where m
i
and n
i
are the
n um b ers of terms and clusters in database i i The termcluster
matrix in the directory of services is the sup erset of those matrices in the
three serv ers where m and n are the n um b ers of its total terms and clusters
and k is the n um b er of orthogonal indexing dimensions used bySVD
w e generate three pseudo do cumen ts d
d
and d
d
fdistribute d network op er ations systems g d
fmanagement monitoring network networks public g d
fapplic ations datab ases distribute d network op er ating systems g where d
is the union of terms in category headings C and C and d
and d
are the unions of
terms in the subcategory headings under C and C resp ectiv ely Similarly w e can generate
a termpseudodo cumen t matrix to represen t the whole taxonom y By merging the sup erset termcluster matrix with the termpseudodo cumen t matrix in the
directory of services and applying SVD on them w e can create a customized directory of services
whic h con tains the decomp osed term and cluster v ectors in fa v or of the terminology used b y the
merged taxonom y Belo w w e dene the mer ge op eration and use Example to demonstrate the
customization Figure sho ws our prop osed system with m ultiple taxonomies
Denition Let A and B be t w o termdo cumen t matrices where A consists of m
terms ro ws
and n
do cumen ts columns and B consists of m
terms ro ws and n
do cumen ts columns Let
T
A
and T
B
b e the sets of terms in A and B resp ectiv elyIf A mer ges B is equal to C denoted
as A B C then C is a termdo cumen t matrix consisting of m terms ro ws and n do cumen ts
columns where n n
n
and m jT
A
T
B
j The coun ting measure j j giv es the size of the
set
Example Assume p
and p
are t w o pseudo do cumen ts created b y the A CM taxonom y where
p
ft
t
g p
ft
t
g W e generate a termpseudodo cumen t matrix from p
and p
and merge it with the termcluster
matrix in Figure b The result is sho w ed in Figure T o examine the eect of merging a taxonom yw e calculate the Euclidean distance b et w een
cluster c
i
and c
j
denoted as dist c
i
c
j
b efore and after the merging T able and Figure sho w
the distances and dimensional co ordinates of clusters when merging with the A CM taxonom y In T able the distance b et w een clusters c
and c
decreases from to This
c hange is b ecause pseudo do cumen t p
con tains b oth t
and t
whic h increases the correlation
bet w een an y cluster con taining either t
suc has c
or t
suc has c
Similarly the distance
bet w een c
and c
decreases due to p
s con taining t
and t
F or clusters c
and c
they do
not ha v e common terms with the same pseudo do cumen t therefore their correlation decreases ie
dist c
c
increases
F rom the example ab o v e w e can see the distance c hanges reecting the correlations b et w een
do cumen ts or clusters in the seman tic space When merging with a taxonom y those clusters
ha ving common terms with pseudo do cumen ts mo veto w ard eac h other By the denition of cosine
similarit y measure all the clusters within a predened angle with the query are relev anttoit
term-cluster
matrix
db1
(m1 xn1)
term-cluster
matrix
db2
(m2 xn2)
term-cluster
matrix
db3
(m3 xn3)
term
matrix
(m xk)
cluster
matrix
(n xk)
SVD
(k)
term-cluster
matrix
(m xn)
Directory of Services
Server1
Server2
Server3
term
matrix
(m6 xk)
cluster
matrix
(n6 xk)
SVD
(k)
Directory of Services (ACM)
term-pseudo-doc
matrix
ACM
(m4 xn4)
Taxonomy
term
matrix
(m7 xk)
cluster
matrix
(n7 xk)
SVD
(k)
Directory of Services (IEEE)
term-pseudo-doc
matrix
IEEE
(m5 xn5)
Taxonomy
Figure SecondLev el LSI with m ultiple taxonomies where m
i
and n
i
are
the n um b ers of terms and clusters in database i i The term
cluster matrix in the directory of service is the sup erset of those matrices
in the three serv ers where m and n are the n um b ers of its total terms and
clusters and k is the n um b er of orthogonal indexing dimensions used b y
SVD In the t w o customized directory of services m
n
and m
n
are
the n um b ers of terms clusters in the A CM and IEEE taxonomies m
n
and m
n
are those n um b ers after merging with the sup erset termcluster
matrix

c
c
c
t
t
t
t
t
t
t
t
p
p
t
t
t
t
c
c
c
p
p
t
t
t
t
t
t
t
t
Figure The termcluster matrix merges with the A CM termpseudo
do cumen t matrix
cluster b efore after c hange
distance merging merging dierence
c
c
c
c
c
c
T able Cluster distances b efore and after merging with the A CM taxon
om yThe c hange dierence column sho ws the distance increasing or decreasing rates

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2-D Plot of Terms and Clusters
Dimension 1
Dimension 2
c1
c2
c3 c1’
c2’
c3’
p1
p2
Figure The clusters in dimensional space when merging with the A CM
taxonom y The A CM pseudo do cumen ts p
and p
are sho wn as The
clusters b efore the merging are represen ted b y c
i
i sho wn as and link ed b y dashed lines The clusters after the merging are represen ted
b y c
i
i sho wn as and link ed b y solid lines

Based on this those clusters close to eac h other are lik ely to b e hit b y the same queryTh us b y
generating pseudo do cumen ts appropriatelyw e can alleviate the p olysem y problem Our metho d
can also b e applied to a small information system where eac h user preselects do cumen ts of in terests
as the user prole and merges it with the termdo cumen t or termcluster matrix at the serv er
Then the serv er can matc h do cumen ts and queries based on the terminology that users are familiar
with or frequen tly use
Design of Exp erimen ts
Toev aluate our solutions w e will conduct exp erimen ts on four standard do cumen t collections
MED CISI INSPEC and CA CM where queries and relev an t judgmen ts are a v ailable W e use
SVDP A CK C pac k age to compute SVD on termdo cumen t matrices and measure the precision
and recall ratios for dieren t metho ds
In our exp erimen t eac hdocumen t collection is likea serv ers database W e do not test LSI
on a single serv er In stead w e generate cluster represen tativ es for eac h database collect them at
the directory of services and use LSI to rank relev antserv ers for eac h query Our goal is to giv e
high ranks to the serv ers that con tain the most relev an t do cumen ts
Tov erify the ranking estimated b y the directory of services w e query eac h serv er with the
same set of data and rank them according to their n um b ers of returned do cumen ts W e calculate
the Sp e arman r ankor der c orr elation c o ecient to measure the closeness of the t w o rankings
If they are iden tical it means the directory of services giv es the user a p erfect hin t of selecting
relev antserv ers
T o examine the eect of merging a taxonom yw e generate pseudo do cumen ts from the
A CM taxonom y and merge them with the do cumen ts in the CA CM database Because the
CA CM database consists of do cumen ts in computer science the adding A CM taxonom y should
help in clustering do cumen ts under the same categoryW e exp ect to see higher recall and precision
ratios
Summary
Weha v e prop osed a new metho d based on Deerw esters laten tseman tic indexing to solv e the
v o cabulary problem in In ternet resource disco v ery Wedesignt w o exp erimen ts to ev aluate our
metho d and exp ect to see impro v ed results in b oth synon ym y and p olysem y References
G W F urnas T K Landauer L M Gomez and S T Dumais The v o cabulary problem in
h umansystem comm unication Communic ations of the A CMv ol pp No v em ber
S Deerw ester S T Dumais G W F urnas T K Landauer and R Harshman Indexing b y
laten t seman tic analysis Journal of the A meric an So ciety for Information Scienc ev ol pp Septem ber
K Obraczk a P B Danzig and SH Li In ternet resource disco v ery services Computer v ol pp Septem b er G Salton and M J McGill Intr o duction ToMo dern Information R etrieval McGra wHill
Bo ok Compan y G Salton A utomatic Information Or ganization and R etrieval McGra wHill Bo ok Compan y A Em tage and P Deutsc h Arc hie An electronic directory service for the in ternet in
Pr o c e e dings of the Winter USENIX Confer enc e pp B Kahle and A Medlar An information system for corp orate users Wide Area Information
Serv ers ConneXions The Inter op er ability R ep ortv ol no pp T BernersLee R Cailliau JF Gro and B P ollermann W orldWide W eb The infor
mation univ erse Ele ctr onic Networking R ese ar ch Applic ations and Policyv ol no pp M McCahill The in ternet gopher proto col A distributed serv er information system Con
neXions The Inter op er ability R ep ortv ol pp July P B Danzig SH Li and K Obraczk a Distributed indexing of autonomous in ternet ser
vices Computing Systemsv ol no pp C M Bo wman P B Danzig D R Hardy U Man b er and M F Sc h w artz Harv est A scal
able customizable disco v ery and access system T ec hnical Rep ort CUCS Univ ersit y
of Colorado O A McBry an GENVL and WWWW T o ols for taming the Web in Pr o c e e dings of the
First International WorldWide Web Confer enc eMa y S F oster Ab out the Veronica service No v em b er Electronic bulletin b oard p osting on
the compinfosystemsgopher newsgroup
P Willett Recen t trends in hierarc hic do cumen t clustering A critical review Information
Pr o c essing Management v ol no pp J E Sammet and A Ralston The new computing review classication system nal
v ersion Communic ations of the A CMv ol pp Jan uary IEEE Service Cen ter Insp e c Thesaurus M W Berry T Do G W OBrien V Krishna and S V aradhan SVDP A CKCv ersion
users guide T ec hnical Rep ort CS Univ ersityof T ennessee Octob er M Kendall and J D Gibb ons R ank Corr elation Metho ds Edw ard Arnold fth ed

Abstract (if available)

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

USC Computer Science Technical Reports, no. 609 (1995)

PDF

USC Computer Science Technical Reports, no. 579 (1994)

PDF

USC Computer Science Technical Reports, no. 637 (1996)

PDF

USC Computer Science Technical Reports, no. 636 (1996)

PDF

USC Computer Science Technical Reports, no. 632 (1996)

PDF

USC Computer Science Technical Reports, no. 611 (1995)

PDF

USC Computer Science Technical Reports, no. 588 (1994)

PDF

USC Computer Science Technical Reports, no. 662 (1997)

PDF

USC Computer Science Technical Reports, no. 652 (1997)

PDF

USC Computer Science Technical Reports, no. 495 (1991)

PDF

USC Computer Science Technical Reports, no. 717 (1999)

PDF

USC Computer Science Technical Reports, no. 780 (2002)

PDF

USC Computer Science Technical Reports, no. 772 (2002)

PDF

USC Computer Science Technical Reports, no. 586 (1994)

PDF

USC Computer Science Technical Reports, no. 595 (1994)

PDF

USC Computer Science Technical Reports, no. 770 (2002)

PDF

USC Computer Science Technical Reports, no. 718 (1999)

PDF

USC Computer Science Technical Reports, no. 704 (1999)

PDF

USC Computer Science Technical Reports, no. 867 (2005)

PDF

USC Computer Science Technical Reports, no. 574 (1994)

Description

Shih-Hao Li and Peter B. Danzig. "Vocabulary problem in internet resource discovery." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 594 (1994).

Asset Metadata

Creator Danzig, Peter B. (author), Li, Shih-Hao (author)

Core Title USC Computer Science Technical Reports, no. 594 (1994)

Alternative Title Vocabulary problem in internet resource discovery (title)

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)

Tag OAI-PMH Harvest

Format 13 pages (extent), technical reports (aat)

Language English

Unique identifier UC16269944

Identifier 94-594 Vocabulary Problem in Internet Resource Discovery (filename)

Legacy Identifier usc-cstr-94-594

Format 13 pages (extent),technical reports (aat)

Rights Department of Computer Science (University of Southern California) and the author(s).

Internet Media Type application/pdf

Source 20180426-rozan-cstechreports-shoaf (batch), Computer Science Technical Report Archive (collection), University of Southern California. Department of Computer Science. Technical Reports (series)

Access Conditions The author(s) retain rights to their work according to U.S. copyright law. Electronic access is being provided by the USC Libraries, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Repository Email csdept@usc.edu

Inherited Values

Title Computer Science Technical Report Archive

Description Archive of computer science technical reports published by the USC Department of Computer Science from 1991 - 2017.

Coverage Temporal 1991/2017

Repository Email csdept@usc.edu

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)