I/P Engine, Inc. v. AOL, Inc. et al

Filing 129

Claim Construction Brief Plaintiff I/P Engine, Inc.'s Opening Claim Construction Brief filed by I/P Engine, Inc.. (Attachments: # 1 Exhibit 1, # 2 Exhibit 2, # 3 Exhibit 3, # 4 Exhibit 4, # 5 Exhibit 5, # 6 Exhibit 6, # 7 Exhibit 7, # 8 Exhibit 8, # 9 Exhibit 9, # 10 Exhibit 10)(Sherwood, Jeffrey)

Exhibit 2 

US006775664H2

INFORMATION FILTER SYSTEM AND
METHOD FOR INTEGRATED CONTENTBASED AND COLLABORATIVE/ADAPTIVE
FEEDBACK QUERIES
5,649,186
5,734,893 A
5.842.199 A

Inventors:

Inc., Waltham, MA
to any disclaimer, the term of this
extended or
under
467

Oct.

A *
*

5,884,282 A

Amram el al. ................ .
Hender;,on et al.
395/149
Yaksich ct al.
Yaksich ct al.
395/149

Farry
Miller et al.

707/4
707/2

continued on next
()'I'IIER PUIIIJ('AI'IONS
Persin, "Document
Fast
of the Seventeenth Annual International ACM-SIGIR Conterence on Research and
in Information
Retrieval, Ju1. 6, 1994 pp. 339-348.

No.: 10/045,198
Filed:

*

7/1996
8/1996
10/1996
10/1996
3/1997
7/1997
3/1998
11/1998
2/1999
3/1999

2001

Prior Publication Data
US 2002/0120609

continued on next

Aug.

Related U.S.

Data

Examiner-John E. Breene
Assistant Examiner-Kuen S. Lu
Finn-Testa, Ilurwitz

Thiheault,

ABSTRACT

Int. CI?
U.S. CI.
Field of Search

G06F 17/30
707/3; 707/5; 707/1; 707/10;
707/2
707/1, 3, 5, 10,
707/2

References Cited

u.s. PATENT DOCUMENTS
5/1991
ct al.
364/192
5/1992 Tirfing ct al.
707/5
9/1993 Haule ......................... .
11/1995 Kawaguchi el al.

system
search
to make one-shot or demand searches for information
whleh
at least threshold matches to user
The search
system also
collaborative/content-based filter to make
searches for information entities which match
and are ranked and stored
accessible, system wires ....Ull .... "fJ'Ull'Ull.l).',
A user feedback
collaborative
for
content
data in the
the collahorative/content-hased I11ter.
query
processor determines whether a demand search or WIre
search is made for an
query.

38 Claims, 10

16

Sheets

u.s. PATENT DOCUMENTS

to Personalized Information
Institute of
707/5

6,029,141
6,029,161
6,078,916 A
6,182,068 BI
131
6,314,420 131

1/2000
212000
2/2000
6/2000
1/2001
10/2001
1112001

Culli,s
705/27

707/1
Culli,,;,
Lang
aL
Lang ct aL

Dumais et aI.,
Access
Textual Information", In
Conference on Human Factors
pp.

707/3

OTHER PUBIJCATIONS
Knowles, "Software Bots filter Web data;
Delivers Customized Information for BBN's PIN", PC
Week
nl7
1995 pp. 12-113.
Architecture for
Filterl'f()Cc:ecllll~~s of the Conference on ComWork, Oct. 1994 pp. 175-186.
"""'U'''"',,, et aI.,
Collaborative
to Weave an
Information
, Communications of the ACM vol.
35, No. 12, Dec. 1992 pp. 61-70.

1991, pp. 1-12.

Fischer et aI., "Information Access in
Structured Information
, Human Factors in
CIII '91 Conference
ACM, 1991
8 pp.
cited

examlller

1

2

3

4

~210

215

5

~310

~315

6

7

8

IN

i

----~-

30C

FEEDBACK PROCESSORFEEDBACK DATA RELEVANT
TO CURRENT INFORMON

50C

VII1EB
SITE

A
38C

40C
:,c

0:::

~

43C

ADAPTIVE
FEEDBACK DATA

w

z

AGENT
MIND
MOLD

42C1

U:II~:::..u.JlJ:;..=t;J.LI;.;.n;.d:l..JI..!L\..u::L"'l

42C

42C2
VII1RE
PR OFI LE S

DEMAND
PR OFI LES

\C

CONTENT-BASED
PROFILE

VII1EB
SITE

M
48C

DEMAND
MODE
RATING

_________ ..J

WIRE ROTAllON TO USERS RELEVANT INFORMONS LIST
72C
ADAPllVE FEEDBACK DATA FOR OTHER USER CONTENT PR

SPIDERCONllNOUS
INFORMON v6BC
ACQUISlllON

ADAPllVE-,-,=-"""",,,,,",,,,,,-,-,,""-'..!..Q.!..~~~~

COLLABORAII VEl
CONTENT-BASED
FILTER
SllRUCTURE

FEEDBACK PROCESSORFEEDBACK DATA FOR
ICONSIDERED INFORMONS

U)I

I
66C

MEMORY

(!)

z

i=

U)

z

i=

«
0::

(!)

~

Z

~

(!)

Z

LL

~

LL

~

SEARCH

U)

X

0::

0

ill

60C

INFORMON
ACQUISIllON
ON DEMAND

QUERY
ROCESSOR

DEMAND
SEARCH
OTHER
USER
STAllONS

I~

Iill
,Z

i=

0
:2

0

78CM

0::

«
0::

0::

62C

74C

::.::
0::

ill

Z

0

I

U)

BOC

•

1

1

2

INFORMATION FILTER SYSTEM AND
METHOD FOR INTEGRATED CONTENTBASED AND COLlABORATIVE/ADAPTIVE
FEEDBACK QUERIES

and informons obtained from the network are '"U"'f""'vU
the
for
and
A
operscans the network to find informons which are
fJ"j ... ',M ... " for
the individual user's

This is continuation of U.S. Ser. No. 09/204,149 filed
Dec. 3, 1998 now U.S. Pat. No. 6,314,420, which
a
continuation in part of u.s. Ser. No. 08/627,436, filed
4, 1996 now U.S. Pat.
5,867,799, the disclosures of
are
reference herein.
This
of U.S. Ser.
No. 09/195,708 filed
6,308,175, which is continuation in part
4, 996 now U.S. Pat. No. 5,867,794,
08/627,436, filed
which are
referthe disclosures
ence herein.

The
filter system compares received informons
to the individual user's query
combined
collaborative
and ranks,
order of value, informons
found to be relevant. The system maintains the ranked
informons in stored list from which the individual user can
select any listed informon for consideration.
As the system continues to feed the individual user's
, the stored relevant informon list
factors
return of new and more relevant
informons,
in the user's query, feedback evaluations
the user for considered informons, and upuauu",.,
in collaborative feedhack data. Received informons are
nn,{'{'",,,y, for other users' wires established in the
information filter system. Thus, the
I1lter system
continued
compares network informons
lind
informons
various users' wires over the
course time, whereas conventional search
initiate
to
individual user's
and
a search in
content-hased
compare the query
accessed
to llnd
informons
a limited, short-term search time
The present invention is directed to an information prosystem
for use
internet
network searches for information
with collaborative feedback
entities relevant to user
I1lter
and content-hased data and

j()

BACKGROUND OF THE INVENTION

responses to user
In the
of the internet, a countless number
informons are available for
from any of least
thousands of sites for consideration
a user at the user's
other web
enters a
informons relevant
a
Thereafter, the search site
"""H11lH'", system and a content-

30

search results.
SUMMARY OF TIlE INVENTION

query.
slve a...,c.U.1UJ.11c:,
the query. More
result in shorter
may also he
return lists. In some
structured to find web
have stored
informons
the entered query.
Collaborative data can
made available to assist m
in form on
when a user
downloads an
informon, considers and evaluates it, and returns
to the
search site as a
of the value of the considered
informon to the user.
In
which is parent to this
Ser. No. 08/627,436,
4, 1996, and

Another system controls the
system to I1lter for one of a wire response and a demand
response and to return the one response to the user. The
system combines
feedback
system with content
data in
of the informons for inclusion in at least a wire
response to the query.
50

BRIEF DESCRIPTION OF TilE DRAWINGS

present invention.
FIG. 2

another

embodiment of an
60

informons
filter structure in this
and collaborative
informons received from various sites in the internet or other
network. In
user enters a query and a correis estahlished, i.e., the query is
m
and
over time,
storage on a content

information

65

information
tion.
FIG. 5 is
an information
invention.

"'-'_''''CUUI", to the present inven-

to the present inven-

to the present

3
FI G.
and
FIG.
FIG.
feature
FIG.

4
model

6 is an illustration of
with associated pn;OlCIClrs.
7 is an illustration of a HHHU,",U<H
8 is a
of the invention;
9 is functional block

j()

user
structure for responses.
DETAILED DESCRIPTION OF TilE
EMBODIMENTS

The invention herein is
with an
apparatus and method for information
in computer
system
a data stream from a computer network, in
which entities of information relevant
the user, or
"informons," are extracted from the data stream
content-based and collaborative
The information
term
the sense
operates on a
and is both interactive
in
structure and method. It is interactive in that communication
is
bi -directional at each level
the filter. It is
distributed in that all or
include a
structure or method, a
method, or a combination
structures and method.
used herein, the term "informon"
information
of
or actual interest to a par"l.:.v"""vc,,,

30

the member client, can increase the
for others of
factor, A's
the user's member clients. Each of the clicnts of one user can
be associated
the individual clients of other users
of the
clients have similar
insofar
the
attributes. A
is
group of clients, called
member clients, that have similar member client
i.e., that share a subset of attributes or interests. In
the subset of shared attributes forms the
for
and is
norms, or common client attributes.
The "relevance" of a
informon
describes how well it satisfies the user's information need.
The more relevant an informon is to a user, the
the
content. The less relevant the inform on, the
the "noise" content.
the notion of what is relevant
a
user can vary
time and with context, and the
informon limited
user can nnd the relevance of
to
Because a user's interests
to the data stream, it
to use
to track the user's current interests
time.
stamp
method
apparatus track the evolution of "relevance"
a user and the communities of which the user a member.
In
information
is the process
the information that a users wishes
see, i.e., informons,
from a
Content -based
is
process
features from the informon,
e.g., the text of a
determine the infor1llon's
on the other hand, is the
relevance. Collaborative
process of
informons, e.g., documents,
deterwhat informons other users with similar interests
needs found to
relevant.
includes filter

m 40 tive content

Also, informons
the aforementioned
Furthermore, an informon can
terned data, such as data file
repreand can be a combination of any of the
sentation of
of the
in
entities.
informons, may be included in an
relevant to user, and is not within
an informon may
the total data stream may
" Therefore,
informato other types
"v'H1:''''~'' to separate the
from the
"noise."
Also as used herein, the term "user" is
individual in
communication with the network. Because an individual
user can be interested in

and
used herein, the term "content-based filter" means
which content data, such
words, is used in
the
process. In a collaborative mter, other user data
the
process. A collaborative
sometimes referred to as a "content" filter since
it
the task of

50

60

group of user
the member client
assoThe present invenof one of user's
tion can
the learned
member clients to others of the user's member clients,
of the learned
e.g., the user's
that the
author in one interest area as

65

user.
are some instances herein where the term
"content filter" used as
from a collaborative
filter, it is intended that the term "content filter" mean
"content-based filter." The
to include at least a
serviced
apparatus,
member
filter for each member client
communities. For this
thc
tributed in that each of the
content
collaborative
even on different levels, and even if many f1lters
on a
level. The
individual user to be a
communi ties,
member clients
tures
the
communities to
over time. Also a member
client has the
III
sudden
lS

no

5

6
<".VHll.HUllll.y

informon, based on the
and member client
the other member clients believe the content of the informon
to be credible, the
will reflect a

]()

and
or
of each recommendation prorelative to other

lJL1 ... 111U''''''"l[J,

of the recommendation. As before, the
value tokens, or
the requester, in return for the
Furthermore, certain embodiments arc
selt -()pl1ffllZJlllg in that some or all of the

for the informon
reflect untrustworthiness. However, the
and declination of
democratic," in the sense
the bias of others'

to be

Member clients can put their
the line," and
refereed
other

informon,
author, the author's
the reviewers,
and the like, and can be fed back to disellssion pa:rtlc11Ja!ltS.
reviewers, and observers to monitor the responses of others
to the debate. The
those
member clients with top ~"~'''.Ul''''
nities may be used to establish those member clients as

coneach

and recommendation services.
most
related
status as sllch, member clients from other
also can
in the consultation or
recommendation process.
In one embodiment the consultation service,
can be
include consultation
With this feature, a member client can transmit an informon
to the network with a
for
on an issue, for
Other member

tokens, or "info
requester, in return for useful I',U>L',"w",,,
one embodiment of an on-line recommendation
and
recommember clients recommendations
on matters
diverse as local auto mechanics and worldclass medieval armor refurbishers. In this embodiment, the
requester can transmit the informon to the network
the request for recommendation. Other member clients can
to the requester with informons
recommendations or disrecommendations, advice,
with the consultation service, the informons of the

40

value
or
neural networks attempt to reduce the size of the space while
,~,.aHHU.'" the information contained in the
However, the SVD can lose information
the transformation and may
methods that arc
include the TF-IDF
and the MDL
FIG. 1 illustrates one embodiment of an information
apparatus 1 structured for search
in accordance with the invention as described subseherc1nin connection with FIGS. 8 and 9. In
a data stream is
network 3, which can be
internetwork. A skilled artisan would
that
be used

60

65

apparatus 1
for a
users. One
for apparatus 1, for
that user 5 and similar users may be subscribers to
commercial information
service, which can be
the owner of computer system 16.
Extraction means 17
be
and receives
data stream 15 from, network 3. Extraction means 17 can

7
and extract raw informons 19 from data stream 15.
thc raw informons 19 has an information contcnt.
Extraction mcans 17
an
least part of the
content
Raw informons
stream for the presence of raw
thosc data cntitics whosc contcnt idcntifics thcm
"in thc
or of
intcrcst to
1.
means 17 can remove
if the informons arrive from
wasted

8
A
response
mcans 21 can bc r-nm"""pn
29 of user 5
first

]()

In one embodiment of the invention herein, it is
that
mcans 33 bc a

user
thc informon contcnt is relcvant to
of which Uscr #1 is a part.
Filter means 21
I11ters raw informons
23 which are
a
informon that,
member client and
dicted to be of
5. Filtcr mcans 21
filtcrs 27a,b and
each
C;"lJ"<'UVC;

In this manner, trends attributablc to individual mcmbcr elients, individual uscrs, and
individual communities in one domain of system 16
be
and influence, similar entities in other
UVH1<HlJ'" \"H~"~"H.'" of agcnt
containcd within sysentities share
tem 16 to the extent that the
include a computer storage means
the

50

information can be stored for later
in
storagc mcans 31, or may bc
to nctwork 3 for
User #2
remote
for
FIG. 2 illustrates another
embodiment of information
50, in computer system 51.
first processor 52, second processors
ratus 50 can
53a,b, third proccssors 64a-d, and a fourth
55,
dIed the desired information
First processor 52
receive a data stream 56 from,
nnlCtO""nr 52 can serve as
H"_v'm""'~ 58 from
49 and
Because of the inconsistencies
infinite individual differences

response 29. User
response
can be active
feedback,
feedback, or a combination. Active feedfor an
back can include the user's numerical
informon, hints, and indices. Hints
include 1i1(e dislil(e
of an author, and informon sourcc and timelincss. Indiccs
can include
agreement with content or author,
humor, or value. Feedback
29
an actual
informon
mcasurc of thc
Pf()p()s(oa informon to the information need
of uscr 5. Such rclevancc fecdback
to
thc
for
the

60

65

III

the
the modes of

connotations,
tion. Mode variations can be even greater between
communities,
interaction and
among communities. Thcrcforc,
that processor 52 create
sentation for each raw informon, thus
informon charactcrization and collaborativc
invariant
tend to facilitate relevant informon
selcction and distribution within and among communitics,

9

tion
diffcrcnccs bctwccn thc
mons. That difference hetween
content of the informon
the
time the user reviewed it and the content of the
informon in its prcscnt form is thc
of intcrcst.! ()
Proccssors 53a,b, 54a-d may eliminatc thc informon from
furthcr
dircct thc ncw, altcrcd informon thc
the event that nature or extent
the
from the
a "delta" threshold. In
delta thrcshold, onc may
to thc cxtcnt that thc
IS
to the user. The nature of this
can be shared among all
delta threshold can he
sccond proccssors 53a,b,
alone or in concert
processors 54a-d. Indeed,
third processor 54a-d can be the locus for delta
where processors 54a-d
delta
cach mcmbcr client of thc
III
thosc
Second processors 53a,h can I11ter raw informons 58 and
informons 59a,b thcrcfrom.
proccssors 53a, b to
communities, in response to
that are
to the communities.
second processors 53a,b are shown
in FIG. 2, systcm 51 can bc scalcd to support many morc
that
proccssors, and communities. It is
second processors 53a,h extract
59u,h
two-step process. Where
mode-invariant concept rp.1t1"'''Pni
53a,b can
dctailed ,,"UlHlJ.lUllll)

30

artisan
processors 52-55 could be combined ,.uu~"uuuUJ
actual number of processors used in the apparatus 50 could
be less than,
than, that illustrated in FIG. 2, For
in onc cmbodimcnt of thc prcscnt invcntion, first
processor 52 can be in
workstation,
53-55
in additional
with
systcms. Suitable
systems can include those based upon the Intel® PentiumIn fact, thc
of
Pro ™
invention allows
extensive
apparatus 50, in which the number of users, and the commumtlcs
may bc
suitable processors, As described in the context of FIG, 1,
the interrelation
the several
and respective I11ters
trends attrihutable to individual member
clients, individual users, and individual communities in one
and influcncc,
domain of systcm 51 to bc
51 to the extent
domains share
common attrihutes.
The above descrihed
accordance with
a mcthod 100 for
HHCVLmaUU'H

54a-d can receive

memher clients
member clients in
each of users 64a.b can maintain interests in each
of the communities serviced
second
53a,b, and cach
61b,c and 61a,d,
Each memher client 63a-d nr"""'ip~
vnJV'Js(;O mcmbcr client
thc mcmbcr clicnt fccdback 65a-d,
least one of the

HH.V'H""','

collaboration
48a,b, and 62a-d.

iJHI\-",,,,.)l

60

65

is
consumer
ence criteria relative to the communities of which the user is

that coherent
raw informons,
into
words, called tokens. Tokens include punctuation and other
that may
part of the
nno [prrF,d

pVlvl.1U,'"

j()

memher client in a
second
criteria. Method 200 also includes the steps
informons
from a data stream and
from raw inforare correlated with
mons.
informons
one or more of the common client attributes of
and of the member client attributes of the
client to
the

counts for the document is created.
This vector is the size of the total
with zeros for
tokens not
this type of
. n" m n , , ' c modeL While

the
as next step it
invention
the tokens be left in their unstemmed
In this form, the tokens arc amenable
classified into mode-invariant concept components.

e.g.,

can
collection, M, of informons.
conceptfor each
for each
a list of
Ms that is stored for each concept and that can he
index-searched. Each A that is determined to be a poor fit for
M is eliminated from further
Once

stream
with the
informons in the data stream.
informons can include
an"nTHlP content filter
the

50

stage.
each user's pv""Jll,al
member client V of M, for additional
on the memher client's
Each
I1ts V's interests
is selected for V's
informon, or "V-Zine, collection, Z.
informons
This userarc eliminated from
in Z
level stage of
and selection may be pertc)rnled on
centralized server site or on the user's computer.
informons are
to user U
Next, the
for review. Vser V reads and rates each selected
Z
The feedback from V can consist
int,erestillg" V found A to
well as
or have no
Feedback: Did V find the facts,
sources,
and quotes in A to be truthful and credible or not?
Informon Qualities: How docs the user
for
.."r" """"

agreement.

Reason Feedback:
dislike A?
Because of the
Because of the source?
Because
is out-of-date

did the user like or

weather report from 3

information contained in A has been seen
of
information
the
Feedback: Did U liked
within the correct M and Z?
Such multi-faceted feedback

Was it

]()

structure
a database be used,
Ms, that are related
Furthermore, when
U,
concept dues
U to the information lUter can be used to determine
concepts C that describe what U is
if U types in
as a
a.',·""da,.vu Z, then
concepts that have
for the word "basketball" are associated with the new
C seem to
an
new
Z. If no such
that is endowed with the clues U has

operOne embodiment of

concept
word choice used to express concept C. A concept
can include
concept clues
and
concept clues
The PCC and NCC can
combined
a processor to create a measure-of-llt that
be compjprrnmpd threshold.
the combined effect
the pce and Nce exceeds the
threshold,
to concept C;
then informon A can be assumed to be
othervvise it is eliminated from further
PCC is a
set
words,
and other features, such as the source
tend to be
the author, each with an associated
In contrast, NCC
of
in A which contains
and other features, such the source or the author,
each with an associated
that tend
make it more
C. For
the term
that A is contained
about automobiles.
then it
concept
the concept
concept
one or more
means. First, C can be
created
user U. Second,
C can be created
an electronic thesaurus or similar
that can
select from a set of concepts and the
words that can be associated
concept. Third, C can
co-occurrence information that can be
aH,H VLH.'''- the content of an informon. This
that related features of concept tend to
occur more often within the same document than in
Fourth, C can be created
the
of collections, II,
A that have been
one
more U. Combinations of
features that tend to occur
as PCC for the

vv,"""''''''''in M,
f''''U''''"'''''W''U threshold of at least one that can serve
an index
M. Another
of concept
management is that, for each A that does not fall into any of
the one or more M that are indexed
C, the breadth of C
IS
to preserve the IIrst
insofar
if
threshold is exceeded for a
IvU'U.vlll).'; PCC,
the threshold for C.
one embodiment of content an
has been
the
set of C that describe it, and determine which M should
HUVU'''F"

to
continual rm'illTll''''
when a concept C no
documents that
been classified and liked
clients U in a
removed from
there appears
concept C, and articles that have been classified
index
users as
M, and if C does not
the list of M indexed
into M, then M can be added
The
heuristic for

for
and

30

40

60

65

each informon A
for each M, each A
for each U

each of those
process are very similar for
the mcmber client
level, the
subscribed to M,
not
an individual U. Other information about the individual U
can be used
the filter, such what U thinks of what
author writes in other Zs that the user
M prm:essm,g.

of these three types act as
for the next level of the
function, called
Correlated-Feature, ErrorCorrection Units
420. From
420,
functions 425a-d can be
I'ntTlnlltf'ci functions 426a-d, 428a-d of the
two functions are the
426a-d, and the
428a-d. IRPs 426a-d can be
UPs 428a-d, so that the more certain an IRP
its
Each
IRP 429a-d
IS
with other IRPs 429a-d in combination
be
function 427a-d. This combination function 427a-d
additive function to far more
from a
cOlnplex neural network function. The results from this are
the
across all UPs, from
and combined
the
430. Once the eWF
nn,jprrpd that result

for
the

map
function
SFI 405 can include vectors of

Minimum
can be used
to determine when to store or use a more" local" component
of the IRP. As a
the
IRP
two of the above terms:

other referenced informons, and the like. It is
vector exists for
include,
funniness, conviolence content,
content,
scientific merit,
of information content, artistic
entertainment value, trendiness/
to future directions,
agreement.
Each CFECU 420 is a unit that can detect sets of
feature combinations which are p.,,,,·p.rl1](ln" 1Il
author X's articles are
disliked in the
except when X writes about lathes.
When an informon authored
X contains the concept of
CFECU 420 is
to
sent to offset the
because of the
Z.

Token 1
IRP
IRP

30

the form of Structured Feature Informa-

value in the
SFI, below, accounts
B, and C, have
assume three
submitted ]() articles that have been read, and
have been rated as in TABLE
the text of this
In the " ....'~V1H~'''H

40

Token
8.73
fUn

12.89
4.27

11.27
5.04

manner similar to the
actual document vectors can be
then

<'"m,,,,j prJ m

seven issues. First, there
no a
'~i"",cuUb which lIsers
have rated an informon
A, before
a
for a user U, who hasn't
read informon A. Therefore, a model for nn°.r1lil'Tlnn
no matter which subset of
be available, if

that have not yet been rated, and, in
B, C,
a new author D has contributed,
the Author field, that
takes sums of the averages, would
follows:
IRP
sum of
the author so
the author so far in this
all
far in this
the author so far
the author

far in this M

all authors
all authors

a
a
a

a

The purpose of the
sum is to make use of broader,
statistics, when strong statistics
a
an informon
author, within a
Z lIlay not yet
When stronger staare available, the broader terms can be eliminated
smaller
This
scheme is similar to
that used for
CWFs
whole.
Some of the averages may be left out in the actual storage

60

are correlated with the B-Team at the 0.5 level, and are
correlated with user C the 0.9 level. Now, suppose user C
. Based on that C's
reads an article and rates
it is reasonable to
that A's
also
be a "5".
Further, suppose that
B-TealIl reads the
article,
it
methods
be:

65

In
rate the article, it should not aiIect the

17
that other B-Team

and would he:

arc
correlation anlong one another
to the present invention can "",,",,'''',",,,
compensate for such inter-user correlation.
Sixth, information about the
of
known, other than
user's
information can include the present
the users like,
what authors the users like, etc. This information can make
the system more effective when it is used for
stronger associations between
hecause Users and B
M have never
read and rated an informon in common, no
correlation between their likes and dislikes can be made,
based on common
alone. However, users A and B
have both read and liked severalinformons authored
the
same author,
Users
and B each
a
dilTerent Zs. Such information
the inference that there is
between
user A's interests and user B's interests. For the most part,
collaborative
systems can not take advantage
this
Seventh, information about the informon under consideration also known, in addition the
it so far.
For
from
that informon A is about the
, better use can be made of which
users
more relevant in the
of the information in the informon. If user B's
user
D's
l S , but
D when inform on A is
between User B's
to
hetween User Band

Next, the
between the
of the member cllent
and informon concan
A neural network could be
so that the error in
used to learn how to compare
is minimized. Ilowever, the invention can
be embodied with use of the invention can be embodied with
use of
cosine
metric, like that
connection with Unstructured Feature Inforcan be
]()
The method lIsed prc:tel'ablly

method, which would be
the Author Held
rated authors K
may appear as follows:

User

Gi''len
to
ALlthor
K

#111
sample

Gi''len
to
Author
L

#
sample

Gi''len
to
Author
M

30

Further, the
component of the member client
of user A and user B may be
correlation of each author under "nmrl:lri",'m
the
function F of the

40

is used. This function can be:
where p and q
domain of the
has
no
a user, then the function of the zero
of
This is because the fact that the lIser
the author can
some
not
an informon
the user. In this case, the
articles
function H of the
far, because it becomes more
inforhas

it
an informon on the nn'.rllrtr'.rl
second user, B. For "AaHl~W",
read and rated in common a certain number
informons,
the
user
of informon D on the
of informon D for user B can
defined
~vWl'C"'~'.H0. First, there can be
user A and user B
M,~profile(A)xprofile(B)xDocumentProfile(D).

Second, a correlation may
taken between user A's past
and user B' s past
with respect
informons
that are similar to D. This correlation can be taken
E that A and B have rated in
of E to D,

#
sample

and M.
60

Author
K

65

O(not

19
'"'"U'I.H~·A '"'"U'I""'""" informon

and

-continued
Author

nr, .. tprn,rI

L

Il71J

G(allthor)
x G (not author)

~

0.02

be used as the
. Also

II)

which separate users into those
common interests. Furthermore, each
can be
broken into SUlJ-S.UD-1ll
503c-d,
to which users
members. As used herein,
to
502a-e, and
the

can
conend nodes. Users 505a,b can be members of
502a, 502e, if such more
matches
their interests than would
In

one-neuron neural

retrieval would recognize that there arc numerous methods that can
used to
effect
document com-

same

cosme
author, can be handled
them as another token in the vector. However, the token is
to be
factor that is
TF-IDF
Each component of the
combined

manager may be used
of the
501,
the

informainformation up to the
parent node, if such exists;
estimations of the
consensus on the
in forman from the
parent
and
estimations of the
consensus on the
for a
informon for the users that come under

additive function,
function,
for
function,
upon
straints encountered in the "ppH~a,.lV'''.

40

In one embodiment of the present invention, it is
that users be broken into distributed
called "mind-

to the aforementioned
of many subcommunities. These
used to represent dilTerent
that can be maintained

that all users be members of the
the top mind-

separate hierarchies
funniness; valuableness;
violence content; sexual content; probusiness
scientific merit; artistic
dramatic
entertainment value;
U11lex pe:ct,;dlle,;s of information content; trendiness or
agreement.
tance to future directions; and
these
can be
users with a
feedback mechanism and, therefore, these
used to drive separate
hierarchies.
can be used in combinations, if

50

1Il

GO

G5

manager is an end node in the
user's CWF, for
the user's
Function 4 also can
the estimations
received from the
parent node, and
Predictions can
estimated based on
size, standard
to above, users can
one
if
""'U~'p'''H but have
of the informon.
vIews
Also, it
between peer managers who have similar users beneath
them to share estimation information. When a
in from user, it can be
node above that user. It is
decide whether the
node. If the manager estimation would
amount
minimum threshold, then the
above an
manager should
that estimation down to all of its
child-nodes. In the event that the
statistics are
more than another minimum threshold amount,
statistics should be
the manager's
if any, and the process recurses
and
downward in the
to have accuBecause no W''''''P''C'' manager is
an estimation of the
with

of
if none of its child-nodes has any
information yet. The
distributed strategy tends to
reduce the communication needed between processors,
the
tends to be
of
of other nodes, and the other users
arc
minimized. Therefore, as the number of informons and users
the
and f"'~U""U"'"
the sum of the numher of informons and the numherlo
users, rather than the
of the number of informons
and the number of users. In addition, incremental
the
of estimations up
of
hecomes stahle

~A"H'IJ"~

oc,,,,rnnh,,,n violations.
with the B-Team of ten users, the
With the

estimation

Lscr Frs
Raling of
informons About

not

estima tion then can be combined with other estimations to achieve
of how many B-Team members
the desired result,
the article at any
time.
hierarchies can be created
methods. I f the
creation is
of
based on information such as
common interests, or any other information that is known about
them. Where the
is created
the

Author As
Article
Other
\Veighted by Topic

1.4
1.68

1.87

Gardening

Politics

Topics

7

40
20(J

].69
1.84

30

Author
Arlicles
Other Authors

70

In
40

1I1(;rem,;nl ally-ad.lm;tlllg the clusters as new information is
is intended
featurcs
would tend to not rank informon D very
the first CFECU would first find sources
include
the

in order to determine how rclevant the sub-informant is in
estimators also arc
the estimation.
so that
can be
the
most accurate
evaluation can
suh-informant. This type
tage of the two

111

combination

50

appears, and has statistical
with the common characteristics of, for
the same author and
CFECU is created to
that this
to the

60

in f Ofm 0 n,
the estimation. It is
used be

correlation neural network, in
which the neural net finds new connection neural net units
f''''U'''U.UU error. Another method
HH.Ull.HUH that has been rated but
error, and
the

65

can
Predictions of the sub-informants, in a manner similar

been found
error and common characteristics, the common characteris-

or a low
Then, the
for all
the
can
the CFECU should be added to the
If the estimated error reduction greater
threshold level, the CFECU can be added
successful CFECU are discovered for
of
If ]()

When a user makes query for which a wire
exists, wire search results arc
returned instead
search
results. As shown in the
of FIG. 7, a user
The query is
to
block 22C, block 24C
table whether a wire
exists for the new query. If so,
block 26C returns results from the
block 28C commands
With the use of wire search returns, each user can review
the returned results and
feedback data about
rev1cwed documents. Such feedback data is
III
the
used in
informons for the wire.
Therefore, when future user makes
query, the wire will have been lITlnrmlP(1
tion of
users' feedback data.
ments which users rate as
the
can
common document
features which can be used to return more like documents

is
system,
terms, and all new
the network
LJU"U"U"UC"'" may be used to send
new documents to the user.
system structures, an
rm,ln,,,,~rI in which collaborative
demand
or Wlre
basic structure and
if
web sites, user
relevant informons across the web. Content-based
lS
used
the
informons, and
the search results are resented III the form of
list
informons ranked
The present invention combines collaborative
with content-based
and further
of the content-based
results, the invention can be embodied as a search
system in accordance with different basic structures. In the
basic structure, an
collaborative/content-based filter

40

50

on the basis
structure of the invention is

~~I"~'_WH

no wire exists
query. Otherwise, wire search
wire does exist,
collaborative
can be
from the wire filter structure to
results of the demand search from the
In the
for the most common
system. A suitable
is
determine which
are most common,
created for each of these
can be made from time
time
to make wire additions or deletions as warranted.

system structures of the
invention arc
embodied with the usc of a procomputer system.
Collaborative
additional data
other
search results for
individual user for
conducted. The collaborative
data, and/or it can be
data for agent mind
which
disclosed in Serial Number
# LYC
INTEGRATED
FILTER STRUCTURE EMPLOYING SELECTIVELY
SIIARED, CONTENT-BASED PROFILE DATA TO
EVALUATE INFORMATION ENTITIES
A MASSIVE
INFORMATION NETWORK, filed
on Nov. 19, 1998, and
reference.
used. For
types of user
have read
medium,
a document. Further,
time spent
users on each

60

65

can be
of
of other document
connection with FIGS. 1-7.
Feedback
number of
invention
method of
information use. Use methods range in
relative ranks
a set amount

FIG. 9 shows a

b~H~"<U.'L~"

the processor 48C rates informons returned
the
system 46C in a demand search indicated
the reference
character 48C2. Collaborative
data is used in the
informon
proeess in the wire search mode, and if
in the demand search mode, to the extent that
collaborative data is available for the informons in the
search return. Search results are returned to the users 34C
embodiment of the invention
and 36C from the search return processor 48C as shown in
CASE system 30C are
FIG. 9.
demandj()
The invention is
embodied
shown in FIG.
10. A query processor 60C receives
from
individual user 62C and other users 64C and determines whether
wIre
for each entered query.
exists,
the query is routed to collaborative/content-based filter
system 68C
structure 66C lice that of FIGS. 1-7. A

the invention, the wire
exists, and wires exist
entered as
demand search mode, the
to normal search
Otherwise, various schemes can be used for
whether a wire search or a demand search is
every query can call for a wire search,
dem and search
made the
is entered and with wire searches
entries of the same query. A'S another ~A"H.'F"V,
select demand search, or, if .... VHULH'lHlb
is desired, the user may select a wire search.
The I11ter structure 40C

a threshold-level match for content
the top level
multi-level I11ter structure, at least one of which reflects the
Informons which
wires are stored III a
wires to which
memory 72C " ....'AWUHJ.b

a

feedback processor 74C structured like the
system of FIG. 7
collaborative feedback data for
with the content-based data in the measurement
the filter 66C. An informon
30

multi Wire
with informon

as

evaluation, feedback data from users
therewith. These
the

I I' no wire exists

40

with
available informons for return, the
system
a content threshold derived from the content
conducted.
In many instances, it s
system
46C have a
system 46CM which holds an informon
information is stored from infordata
from the network. In this manner,
be
made from the
search and
query from the search
A search return processor 48C
search informons or wire search
content-based filter structure 40C
mode
the latter, and includes
system
system
which is like that of FIG. 6. The
combines content-based
data with collaborative
feedback
from users
feedback processor SOC at least in the wire search mode and, if desired, in
the demand search mode.
In the wire search mode, the processor 48C rates informons on a
are received from the
system 46C as indicated
network 44C
the reference character 48C1. In the demand search mode,

50

query, the query is
a content
is
of informons returned
a
system 78C in a demand search of the network
system 78C can have its own memory
70C. The
considered in connection
the
system 78CM
46C of FIG. 9.
Once
on returned informons, those
Sal.ISlactOlY match to the query
are returned
a list to
user
a search return
processor 80C. The
80C creates a new wire for the
current query for
demand search was made, if
demand search memory 82C indicates that the current query
to
been made over time
sulIicient
as a "common" query for which a wire is
indicated
dashed connector line 80FD, collaborative
feedback data can be, and
into the
demand search prc)cesslng
those
a

scope of the invention. Therefore,
must be
that the illustrated embodiments have been set
for the purposes
and that it should not
the invention as del1ned

ndpe"',,nd

65

trated and described above, what is
the essential
and also what
tion.

the delivered information to determine relevance to the at
one of the first user and the
13. The search system of claim
the feedback
rp':nr,m:p further
information
data.
The search system of claim 13 further .... ,,'",,_0 .. 0."
module to to the "H.V1HHlIRHJ

TABLE 1
Article

3

Raling

Author
A
B
13
B
C
C

5

2

5

]()

K

C

for
user.
17. The search
users CUllH'"',":" at

9

10

2

C

of claim 1 wherein the
one
of

of

TABLE 2
Article Author
IRP (author)

normalized

normalized

Il

12
14

What is claimed is:
1. A search system _~H"~HJH.,,,,
a
system for
for information relevant
user 1Il
users;
a feedback system for rM'~"""a information found to be
users; and
relevant t~ the query
a content-based filter
the information from the lv'~UL"'v"
from the .,v'UH1Hl.".
bined information
relevance
and the first user.
search system of claim I wherein the content-based
I11ters information on
basis.
search system of claim 1 wherein the HlJ.L'liH,1l1'.'1l
an informon.
4. The search
of claim 1 wherein the filtered
information relevant to at least one of the first user and the
query
to
future
the first user.
5. The search system of claim
the filtered
information is an advertisement.
6. The search system of claim 1 further VU"'fJU"H1F'. an
information
system for
the I11tered information to the first us';r.
7. The search system of claim 1 further
information
feedback communication means for
to at least one of the other users.
8. The search system
7 wherein the information
delivered to the at least one of the other users further
the filtered information.
9. The search system
7 wherein the information
least one of the other users further
delivered to the
feedback
claim 9 wherein the information
found to be relevant to the
query
response to the feedback
The search system of claim 7 further
monitor for
time spent
the at
other users
the
information.
12. The search system of claim II wherein the contentbased filter system uses the measured time spent av'"",,,""".".

L8. The search system of claim 17 wherein users in the
of users are
into at
one distributed

30

The search system of claim 17 wherein the m,n,",,'"
of users
. least one of distributed group of users
a hicrarchical structure, a distributed group of users
a
structure, and a distributed group of users
hierarchical structure and
combination of
of claim 1 wherein the contentwherein the contentfeatures from the

40

50

information.
22. The search system of claim 2L wherein the extracted
features
~lIltent data indicative of the relevance
the at least one of the query and the user.
23. The search
of
22 wherein the content
the at least one of the
indicative of
query and the user
elements
information obtained from the information received from the feedback system.
24. The search system of claim 1 wherein the "~'HH.HHM
system further
network
demand search request.
25. The search system of claim 22 wherein the search
auauuyv user feedback data to the contentcomponent for the
information relevant to

first

"""";mll,,, for information relevant to a query associated
with a first user in a
60

65

of users;
relevant to the query

other users;
the
found to
relevant to the
query
other users with the searched information; and
content-based
the combined information for reland the first user.
evance to at
one of the
the step of
27. The method of claim 26
information on a

28. The method of claim 26 further '"U"',""HUI", the step of
the filtered information to the
29. The search system of claim 26 further
the
of
information to at least one of the other

tion delivered to
the

34. The method of claim 33 further NHTlnn'" the
the measured time spent determine relevance
at least one of the first user and the query.
35. The method of claim 32 further

of claim 29 wherein the informaat least one of the other users further
of
the filtered information to

Cnnll1n"1nLf the step
a feedback query to the at least one of the other

users.
32. The method of claim 31 wherein the
of
information found to be relevant to the query further coma feedhack response to the
the step of
query.
33. The method of claim 29 further
the step of
the at least one of the other users

of
the

at least two users in the
]()

37. The method of claim 36 further
the step of
uuuU,""'" users into at least one of a hierarstructure, and a combination of

claim
network in
search for the information relevant
with the first user.

step
to a demand

"~'''-'''''h