Fraud detection using predictive modeling
US005819226A
United States Patent [19]
[11]
Patent Number:
5,819,226
Gopinathan et al.
[45]
Date of Patent:
Oct. 6, 1998
[54]
FRAUD DETECTION USING PREDICTIVE MODELING
[75]
Inventors: Krishna M. Gopinathan; Louis S. Biafore; William M. Ferguson; Michael A. Lazarus, all of San Diego; Anu K. Pathria, Oakland; Allen J ost, San Diego, all of Calif.
OTHER PUBLICATIONS
Electric Academy Electric PoWer Technology Institute Data PE—89—33, “Analysis of Learning Process of Neural Net Work on Security Assessment”, pp. 161—170. Gullo, Karen, “Neural Nets Versus Card Fraud Chase’s Software Learns to Detect Potential Crime” Feb. 2, 1990 American Banker Magazine .
[73] Assignee: HNC Software Inc., San Diego, Calif.
International Search Report, International Application No PCT/US93/08400, mailed Jan. 12, 1994.
[21] Appl. No.: 941,971
—Propagating Errors” Nature v. 323, pp. 533—536 (1986).
Rumelhart, K.E., et al., “Learning Representations by Back .
_
Hecht—Nielsen, R., “Theory of the Backpropagation Neural
[22]
Flled'
Sep' 8’ 1992
NetWor ”, Neural Networks for Perception pp. 65—93
[51]
Int. Cl.6 ................................................. .. G06F 157/00
(1992)
[52] [58]
US. Cl. ................................................................. .. 705/1 Field of Search ................................... .. 364/401, 406,
Weigend, A.S., et al., “Generalization by Weight—Elimina tion With Application to Forecasting”, Advances in Neural
364/408; 395/21, 23; 235/23, 380; 705/35, 1
[56]
References Cited
Primary Examiner—Gail O. Hayes Attorney, Agent, or Firm—FenWick & West LLP
US. PATENT DOCUMENTS 5,025,372 5,416,067
Information Processing Systems 3 pp. 875—882.
[57]
6/1991 Burton et al. ......................... .. 364/406 5/1995 Sloan et al. ........................... .. 235/381
FOREIGN PATENT DOCUMENTS
ABSTRACT
An automated system and method detects fraudulent trans actions using a predictive model such as a neural network to
evaluate individual customer accounts and identify poten
t1ally fraudulent transactions based on learned relationships
0 418 144 A1 0 421 808 A3 0 468 229 A2
3/1991 4/1991 7/1991
European Pat. Off. ........ .. G06F 7/08 European Pat- Off- -------- -- G07F 7/10 European Pat O? GOGF 15/80
among known variables. The system may also output reason codes indicating relative contributions of various variables
A62-74768
4/1987
Japan ................ ..
performance and redevelops the model When performance
A63-184870
7/1988
Japan
. G06F 15/30
A4-113220
4/1992
Japan
.. G01D 3/00
A4- 220758 WO 89/06398
8/1992 7/1989
Japan . G06F 15/18 WIPO ........................... .. G06F 15/30
G06F 15/30
to a particular result. The system periodically monitors its dro S below a p
redetermined level p
'
38 Claims, 21 Drawing Sheets
801
802
805
Model
Transaction
Development
Processing
‘
804
Past Data
Property Data
'
806
Create or update 7
profile
Build Model
Fraud scores and reason
108
codes
Neural Network 1
807
Output
100
Area Data
f
U.S. Patent
106 F'mancla ' 1 D at a
Facility
0a. 6, 1998
Sheet 1 0f 21
5,819,226
102
103
RAM
Data Storage
_\
\ 108
105
101
—D Data Network
D
CPU
104
107
Output Device
Program Storage
100
FIGURE 1
U.S. Patent
—
Oct.6,1998
FALCON
Cutoff Score
Sheet 2 0f 21
5,819,226
Monitor
700 W’ 202
Accounts over cutoff
143 "\_/ 203
Analyst Loading 25 -
Flags
t2‘ 10 W 5 _
O
2
6
10 ‘I4 18 22
Hours
1000— 800 600
'Q
‘\ 207
400- Q 200 O_
I‘?
v
‘*1
— 204
Froud Score 506 5401 01 3767232881 OK
Help
FIGURE 2
k 205
U.S. Patent
0a. 6, 1998
Sheet 3 0f 21
5,819,226
Account Selection Score 886 889 895 898 898 902 902 908 91 1 916 927 932 933 935 943 964 965
Account 5403173602031736 5484743400847434 5467884700678847 4446833257468332 5422023883220238 4419793403197934 4401377501013775 4413703002137030 4406383663063836 4402489633024796 5446281200462812 5403173602031736 5430354200303542 4400021016000210 4412540900125109 5400177963001779 5419472700194727
966§4400613000006130 968 988 993 994
5400004602000046 5403215301032153 5440625003406250 5426836600268366
Evaluate
Restrict
0K
Help
FIGURE 3
I’ 302
U.S. Patent
—
Oct.6,1998
Sheet 4 0f 21
5,819,226
Account Score
Account 4400613000006130
Reasons
Nome Sondro
Score 966
Simpson
~7\403
\J\ 402
1 Suspicious approve/decline pattern 2 Suspicious recent tronsoction rote
3 Suspicious previous day tronsoction octivity Current and Previous 7 days Auth Records Tron Amt Dote Time Avcred CredLim Sic MerchZip 22.10 920320 111856 10.00 1000.00 5399 0.00 29.95 920320 112737 32.00 1000.00 5399 0.00 25.30 920321 235944 61.00 1000.00 5812 0.00 23.04 920322 3624 61.00 1000.00 5331 0.00 54.00 920322 142607 86.00 1000.00 5331 0.00 54.00 920322 142756 86.00 1000.00 5311 0.00 Lost 6 months 10.35 920217 224749 127.00 1000.00 5942 0.00 50.00 920222 230825 685.00 1000.00 5541 0.00 69.27 920223 4446 635.00 1000.00 5812 0.00 10.35 920223 5800 566.00 1000.00 5942 0.00 25.37 920224 202441 556.00 1000.00 5499 0.00 254.70 920229 4803 507.00 1000.00 5399 0.00
i 406
OK
Help
i FIGURE 4
407
U.S. Patent
0a. 6, 1998
—
Sheet 5 0f 21
5,819,226
Cordholder Info
502
Account 4400613000006130 —\__/(/ Nomel
Sandro
Simpson
Nome2
Joseph
Simpson
_I 503 504
Best time to coll _ Phone
Diol
7 — 10 pm
*1’
numbers
Home
612-345-6328
_
Diol Work 1 612—635—2348
L505
Diol Work 2 6l2—325—6723 _ Address
Addrl
915 No. Arlington Heights R0
‘1507
Addr2 City
Minneapolis
State
MN
OK
\
.
1
Qecision
‘Id 507
Zip 55402
Help
FIGURE 5
505
U.S. Patent
Oct. 6, 1998
—
Sheet 6 0f 21
5,819,226
Decision
|:| No contact; all phones II No contact; msg. left
|:| Customer leri?ed chorge(s) @ Customer Qenied chorge(s) |:| Customer _u_nsure of ohorge(s)
I1 Desk gpprovol
Comments
\602
Customer Denied all charges on 3/22/92
L503
OK
Cancel
Help
\
i604 FIGURE 6
\\
507
U.S. Patent
Oct. 6, 1998
Sheet 7 0f 21
701
Train network model
using past data
i
702
Store network model
i
703
Obtain data for current transaction
l
704
Apply network model to current
transaction
i
705
Output results
FIGURE 7
5,819,226
U.S. Patent
0a. 6, 1998
Sheet 9 0f 21
5,819,226
I
903
Transfer function 901
FIGURE 9
U.S. Patent
Oct.6,1998
Sheet 10 0f 21
FIGURE 10
5,819,226
U.S. Patent
Oct.6,1998
Sheet 11 0f 21
5,819,226
w23iu5w
m : Saw5w
m9E%w.m k
Po:
N628
mo:
mwLhax
2:.
wm gop
%
m1; >tH2Dou9m5
$358?
or: >#5m08“w
$32. ,
m : 81 m
PM: NSHGauUQwmE
ow:
mo:
m7dmcw ow:
wmuoacw
Z:
w@E3o82
526i
U.S. Patent
Oct. 6, 1998
Sheet 12 0f 21
1 202
Read past transaction ‘ database
1
1203
Read customer
database
l
1204
Generate new profile record
1
1205
Save new profile record in profile
database
FIGURE 12
More accounts?
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 13 0f 21
1301
Start
1302
Read past transaction database
1 1 303
Read customer
database
1 1 304
Read record from
profile database
1 1 305
Generate updated profile record
1 1 306
Save updated profile record in profile database
FIGURE 13
More accounts?
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 14 0f 21
FIGURE 14 1 402
Record all transactions
throughout the day
l 1403 Obtain current transaction data
+ 1404
Obtain past transaction data, customer data, and
profile data
l 1405
Apply data to neural network
+ 1406 Obtain fraud score
from neural network
Is fraud score > threshold?
1408
Flag account
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 15 0f 21
FIGURE 15
1502
Merchant calls for authorization
l 1503
If account not already
flagged, authorize transaction
l 1504
Obtain current transaction data
+ 1505
Obtain past transaction data, customer data, and
profile data
1 1506
Apply data to neural network
+ 1507 Obtain fraud score
from neural network
Is fraud score > threshold?
1509
Flag account
5,819,226
U.S. Patent
Oct.6,1998
Sheet 16 0f 21
5,819,226
1601
FIGURE 16
Start
1602
Merchant calls for authorization
1 1603
Obtain current transaction data
+ 1604
Obtain profile data
1 1605
Apply data to neural network
t 1 606 Obtain fraud score
from neural network
Is fraud score > threshold?
1 608
1 609
Send signal to authorization system
Send signal to authorization system
indicating high fraud
indicating low fraud
score, and flag
score
account
1 61 0
——> Update profile data 4—
U.S. Patent
Oct.6,1998
Sheet 17 0f 21
1701
Start
1702
CINITNET
1703
CSCORE
4—
More accounts?
1705
CFREENET
FIGURE 17
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 18 0f 21
1801
Start
1802
Obtain current transaction data
1 1803
Obtain data on other transactions within
the last 7 days
1 1 804
Obtain record from
profile database
1 1805
Obtain customer data
1 1806
Generate fraud-related variables
1 1807
Run DeployNet
1 1808
Output score and reason codes
1809
End
FIGURE 18
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 19 0f 21
1901
Start
1902
Scale fraud-related variables
l
1 903
Initialize input layer of neural network
i
1904
Iterate network to generate score and reason codes
FIGURE 19
5,819,226
United States Patent [19]
[11]
Patent Number:
5,819,226
Gopinathan et al.
[45]
Date of Patent:
Oct. 6, 1998
[54]
FRAUD DETECTION USING PREDICTIVE MODELING
[75]
Inventors: Krishna M. Gopinathan; Louis S. Biafore; William M. Ferguson; Michael A. Lazarus, all of San Diego; Anu K. Pathria, Oakland; Allen J ost, San Diego, all of Calif.
OTHER PUBLICATIONS
Electric Academy Electric PoWer Technology Institute Data PE—89—33, “Analysis of Learning Process of Neural Net Work on Security Assessment”, pp. 161—170. Gullo, Karen, “Neural Nets Versus Card Fraud Chase’s Software Learns to Detect Potential Crime” Feb. 2, 1990 American Banker Magazine .
[73] Assignee: HNC Software Inc., San Diego, Calif.
International Search Report, International Application No PCT/US93/08400, mailed Jan. 12, 1994.
[21] Appl. No.: 941,971
—Propagating Errors” Nature v. 323, pp. 533—536 (1986).
Rumelhart, K.E., et al., “Learning Representations by Back .
_
Hecht—Nielsen, R., “Theory of the Backpropagation Neural
[22]
Flled'
Sep' 8’ 1992
NetWor ”, Neural Networks for Perception pp. 65—93
[51]
Int. Cl.6 ................................................. .. G06F 157/00
(1992)
[52] [58]
US. Cl. ................................................................. .. 705/1 Field of Search ................................... .. 364/401, 406,
Weigend, A.S., et al., “Generalization by Weight—Elimina tion With Application to Forecasting”, Advances in Neural
364/408; 395/21, 23; 235/23, 380; 705/35, 1
[56]
References Cited
Primary Examiner—Gail O. Hayes Attorney, Agent, or Firm—FenWick & West LLP
US. PATENT DOCUMENTS 5,025,372 5,416,067
Information Processing Systems 3 pp. 875—882.
[57]
6/1991 Burton et al. ......................... .. 364/406 5/1995 Sloan et al. ........................... .. 235/381
FOREIGN PATENT DOCUMENTS
ABSTRACT
An automated system and method detects fraudulent trans actions using a predictive model such as a neural network to
evaluate individual customer accounts and identify poten
t1ally fraudulent transactions based on learned relationships
0 418 144 A1 0 421 808 A3 0 468 229 A2
3/1991 4/1991 7/1991
European Pat. Off. ........ .. G06F 7/08 European Pat- Off- -------- -- G07F 7/10 European Pat O? GOGF 15/80
among known variables. The system may also output reason codes indicating relative contributions of various variables
A62-74768
4/1987
Japan ................ ..
performance and redevelops the model When performance
A63-184870
7/1988
Japan
. G06F 15/30
A4-113220
4/1992
Japan
.. G01D 3/00
A4- 220758 WO 89/06398
8/1992 7/1989
Japan . G06F 15/18 WIPO ........................... .. G06F 15/30
G06F 15/30
to a particular result. The system periodically monitors its dro S below a p
redetermined level p
'
38 Claims, 21 Drawing Sheets
801
802
805
Model
Transaction
Development
Processing
‘
804
Past Data
Property Data
'
806
Create or update 7
profile
Build Model
Fraud scores and reason
108
codes
Neural Network 1
807
Output
100
Area Data
f
U.S. Patent
106 F'mancla ' 1 D at a
Facility
0a. 6, 1998
Sheet 1 0f 21
5,819,226
102
103
RAM
Data Storage
_\
\ 108
105
101
—D Data Network
D
CPU
104
107
Output Device
Program Storage
100
FIGURE 1
U.S. Patent
—
Oct.6,1998
FALCON
Cutoff Score
Sheet 2 0f 21
5,819,226
Monitor
700 W’ 202
Accounts over cutoff
143 "\_/ 203
Analyst Loading 25 -
Flags
t2‘ 10 W 5 _
O
2
6
10 ‘I4 18 22
Hours
1000— 800 600
'Q
‘\ 207
400- Q 200 O_
I‘?
v
‘*1
— 204
Froud Score 506 5401 01 3767232881 OK
Help
FIGURE 2
k 205
U.S. Patent
0a. 6, 1998
Sheet 3 0f 21
5,819,226
Account Selection Score 886 889 895 898 898 902 902 908 91 1 916 927 932 933 935 943 964 965
Account 5403173602031736 5484743400847434 5467884700678847 4446833257468332 5422023883220238 4419793403197934 4401377501013775 4413703002137030 4406383663063836 4402489633024796 5446281200462812 5403173602031736 5430354200303542 4400021016000210 4412540900125109 5400177963001779 5419472700194727
966§4400613000006130 968 988 993 994
5400004602000046 5403215301032153 5440625003406250 5426836600268366
Evaluate
Restrict
0K
Help
FIGURE 3
I’ 302
U.S. Patent
—
Oct.6,1998
Sheet 4 0f 21
5,819,226
Account Score
Account 4400613000006130
Reasons
Nome Sondro
Score 966
Simpson
~7\403
\J\ 402
1 Suspicious approve/decline pattern 2 Suspicious recent tronsoction rote
3 Suspicious previous day tronsoction octivity Current and Previous 7 days Auth Records Tron Amt Dote Time Avcred CredLim Sic MerchZip 22.10 920320 111856 10.00 1000.00 5399 0.00 29.95 920320 112737 32.00 1000.00 5399 0.00 25.30 920321 235944 61.00 1000.00 5812 0.00 23.04 920322 3624 61.00 1000.00 5331 0.00 54.00 920322 142607 86.00 1000.00 5331 0.00 54.00 920322 142756 86.00 1000.00 5311 0.00 Lost 6 months 10.35 920217 224749 127.00 1000.00 5942 0.00 50.00 920222 230825 685.00 1000.00 5541 0.00 69.27 920223 4446 635.00 1000.00 5812 0.00 10.35 920223 5800 566.00 1000.00 5942 0.00 25.37 920224 202441 556.00 1000.00 5499 0.00 254.70 920229 4803 507.00 1000.00 5399 0.00
i 406
OK
Help
i FIGURE 4
407
U.S. Patent
0a. 6, 1998
—
Sheet 5 0f 21
5,819,226
Cordholder Info
502
Account 4400613000006130 —\__/(/ Nomel
Sandro
Simpson
Nome2
Joseph
Simpson
_I 503 504
Best time to coll _ Phone
Diol
7 — 10 pm
*1’
numbers
Home
612-345-6328
_
Diol Work 1 612—635—2348
L505
Diol Work 2 6l2—325—6723 _ Address
Addrl
915 No. Arlington Heights R0
‘1507
Addr2 City
Minneapolis
State
MN
OK
\
.
1
Qecision
‘Id 507
Zip 55402
Help
FIGURE 5
505
U.S. Patent
Oct. 6, 1998
—
Sheet 6 0f 21
5,819,226
Decision
|:| No contact; all phones II No contact; msg. left
|:| Customer leri?ed chorge(s) @ Customer Qenied chorge(s) |:| Customer _u_nsure of ohorge(s)
I1 Desk gpprovol
Comments
\602
Customer Denied all charges on 3/22/92
L503
OK
Cancel
Help
\
i604 FIGURE 6
\\
507
U.S. Patent
Oct. 6, 1998
Sheet 7 0f 21
701
Train network model
using past data
i
702
Store network model
i
703
Obtain data for current transaction
l
704
Apply network model to current
transaction
i
705
Output results
FIGURE 7
5,819,226
U.S. Patent
0a. 6, 1998
Sheet 9 0f 21
5,819,226
I
903
Transfer function 901
FIGURE 9
U.S. Patent
Oct.6,1998
Sheet 10 0f 21
FIGURE 10
5,819,226
U.S. Patent
Oct.6,1998
Sheet 11 0f 21
5,819,226
w23iu5w
m : Saw5w
m9E%w.m k
Po:
N628
mo:
mwLhax
2:.
wm gop
%
m1; >tH2Dou9m5
$358?
or: >#5m08“w
$32. ,
m : 81 m
PM: NSHGauUQwmE
ow:
mo:
m7dmcw ow:
wmuoacw
Z:
w@E3o82
526i
U.S. Patent
Oct. 6, 1998
Sheet 12 0f 21
1 202
Read past transaction ‘ database
1
1203
Read customer
database
l
1204
Generate new profile record
1
1205
Save new profile record in profile
database
FIGURE 12
More accounts?
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 13 0f 21
1301
Start
1302
Read past transaction database
1 1 303
Read customer
database
1 1 304
Read record from
profile database
1 1 305
Generate updated profile record
1 1 306
Save updated profile record in profile database
FIGURE 13
More accounts?
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 14 0f 21
FIGURE 14 1 402
Record all transactions
throughout the day
l 1403 Obtain current transaction data
+ 1404
Obtain past transaction data, customer data, and
profile data
l 1405
Apply data to neural network
+ 1406 Obtain fraud score
from neural network
Is fraud score > threshold?
1408
Flag account
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 15 0f 21
FIGURE 15
1502
Merchant calls for authorization
l 1503
If account not already
flagged, authorize transaction
l 1504
Obtain current transaction data
+ 1505
Obtain past transaction data, customer data, and
profile data
1 1506
Apply data to neural network
+ 1507 Obtain fraud score
from neural network
Is fraud score > threshold?
1509
Flag account
5,819,226
U.S. Patent
Oct.6,1998
Sheet 16 0f 21
5,819,226
1601
FIGURE 16
Start
1602
Merchant calls for authorization
1 1603
Obtain current transaction data
+ 1604
Obtain profile data
1 1605
Apply data to neural network
t 1 606 Obtain fraud score
from neural network
Is fraud score > threshold?
1 608
1 609
Send signal to authorization system
Send signal to authorization system
indicating high fraud
indicating low fraud
score, and flag
score
account
1 61 0
——> Update profile data 4—
U.S. Patent
Oct.6,1998
Sheet 17 0f 21
1701
Start
1702
CINITNET
1703
CSCORE
4—
More accounts?
1705
CFREENET
FIGURE 17
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 18 0f 21
1801
Start
1802
Obtain current transaction data
1 1803
Obtain data on other transactions within
the last 7 days
1 1 804
Obtain record from
profile database
1 1805
Obtain customer data
1 1806
Generate fraud-related variables
1 1807
Run DeployNet
1 1808
Output score and reason codes
1809
End
FIGURE 18
5,819,226
U.S. Patent
Oct. 6, 1998
Sheet 19 0f 21
1901
Start
1902
Scale fraud-related variables
l
1 903
Initialize input layer of neural network
i
1904
Iterate network to generate score and reason codes
FIGURE 19
5,819,226
