U.S. patent application number 17/676107 was filed with the patent office on 2022-06-02 for matching non-exact addresses.
This patent application is currently assigned to Bottomline Technologies, Inc.. The applicant listed for this patent is Bottomline Technologies, Inc.. Invention is credited to Richard A. Baker, JR., Kaiyu Pan.
Application Number | 20220171753 17/676107 |
Document ID | / |
Family ID | |
Filed Date | 2022-06-02 |
United States Patent
Application |
20220171753 |
Kind Code |
A1 |
Baker, JR.; Richard A. ; et
al. |
June 2, 2022 |
Matching Non-exact Addresses
Abstract
A two-step algorithm for conducting near real time fuzzy
searches of a target on one or more large data sets is described,
where the address and the geolocation are included in the match
criteria. This algorithm includes the simplification of the data by
removing grammatical constructs to bring the target search term
(and the stored database) to their base elements and then perform a
Levenshtein comparison to create a subset of the data set that may
be a match. Location is determined by a geohash comparison of the
latitude and longitude and a Levenshtein comparison of the address
text. Then performing a scoring algorithm while comparing the
target to the subset of the data set to identify any matches.
Inventors: |
Baker, JR.; Richard A.;
(West Newbury, MA) ; Pan; Kaiyu; (Johns Creek,
GA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Bottomline Technologies, Inc. |
Portsmouth |
NH |
US |
|
|
Assignee: |
Bottomline Technologies,
Inc.
Portsmouth
NH
|
Appl. No.: |
17/676107 |
Filed: |
February 18, 2022 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
16655635 |
Oct 17, 2019 |
11269841 |
|
|
17676107 |
|
|
|
|
International
Class: |
G06F 16/22 20060101
G06F016/22; G06F 16/953 20060101 G06F016/953; G06F 16/29 20060101
G06F016/29; G06K 9/62 20060101 G06K009/62; G06F 17/10 20060101
G06F017/10; G06F 16/2458 20060101 G06F016/2458 |
Claims
1. A method for comparing a target address with a base address, the
method comprising: receiving the base address in text format;
looking up a base longitude and a base latitude for the base
address; receiving the target address in text format; looking up a
target longitude and a target latitude for the target address;
subtracting the target longitude from the base longitude, resulting
in a longitude difference; subtracting the target latitude from the
base latitude, resulting in a latitude difference; comparing an
absolute value of the longitude difference to a longitude threshold
value and comparing an absolute value of the latitude difference to
a latitude threshold value; if the longitude difference is less
than or equal to the longitude threshold value and the latitude
difference is less than or equal to the latitude threshold value,
returning an indication of a match; and if the longitude difference
is greater than the longitude threshold value or the latitude
difference is greater than the latitude threshold value,
calculating a score Levenshtein distance between the target address
and the base address in the target, and returning the indication of
the match if the score Levenshtein distance is greater than a
Levenshtein threshold value.
2. The method of claim 1 wherein the longitude threshold value is
the same as the latitude threshold value.
3. The method of claim 1 further comprising storing the base
longitude and the base latitude for the base address.
4. The method of claim 3 wherein the looking up of the base address
and the storing of the base longitude and the base latitude are
done as part of a a batch.
5. The method of claim 1 wherein the base address in text format is
automatically corrected before looking up the base longitude and
the base latitude.
6. The method of claim 1 wherein the target address in text format
is automatically corrected before looking up the target longitude
and the target latitude.
7. A non-transitory computer readable media programmed to: receive
a target address in text format; look up a target location
information for the target address; calculate a target geohash from
the target location information; add a first threshold value to the
target geohash to calculate a geohash high range; subtract the
first threshold value to the target geohash to calculate a geohash
low range; extract all extracted records in a watch list, where the
extracted records contain a watch list geohash value between the
geohash low range and the geohash high range; and identify the
extracted records as a match to the address target.
8. The non-transitory computer readable media of claim 7 further
programmed to loop through the watch list records, if the there are
no extracted records, and calculate a score Levenshtein distance
between the target address and a watch list address in the watch
list, and return each watch list record where the score Levenshtein
distance is greater than a second threshold value.
9. The non-transitory computer readable media of claim 7 wherein
the target information is a longitude value and a latitude
value.
10. The non-transitory computer readable media of claim 9 further
programmed to calculate the geohash by combining bits of the
longitude value with the bits of the latitude value, where even
bits in the geohash are taken for the longitude value and odd bits
in the geohash are taken for the latitude value.
11. The non-transitory computer readable media of claim 7 wherein
the extraction of the extracted records in the watch list involves
a linear search.
12. The non-transitory computer readable media of claim 7 wherein
the watch list is sorted by the watch list geohash value.
13. The non-transitory computer readable media of claim 12 wherein
the extraction of the extracted records in the watch list involves
using the geohash value as a key to locate the watch list
records.
14. An apparatus for determining if a target address is found in a
watch list of addresses, the apparatus comprising: a computing
device with a user interface for accepting the target address in
text format, the computing device configured to convert the target
address into target location information; a storage device
containing the watch list of addresses; and a server, connected to
the storage device, wherein the server operates software
instructions to loop through the addresses in the watch list, and
for each watch list address calculates a difference between the
target address and each watch list address, compares the difference
to a first threshold, indicates a match if the difference is less
than or equal to the first threshold, and if the difference is
greater than the first threshold, calculates a score Levenshtein
distance between the target address and each watch list address,
and indicates the match if the score Levenshtein distance is
greater than a second threshold value.
15. The apparatus of claim 14 wherein the score Levenshtein
distance is calculated before the difference is calculated.
16. The apparatus of claim 14 wherein the difference is calculated
by subtracting a target geohash for the target address from a base
geohash for each watch list address and normalizing the
difference.
17. The apparatus of claim 16 wherein the normalizing of the
difference involves finding a most significant bit of the
difference.
18. The apparatus of claim 14 further comprising an internet,
connected between the computing device and the server.
19. The apparatus of claim 14 wherein the computing device and the
server are the same device.
20. The apparatus of claim 14 further comprising an internet,
connected between the computing device and the storage device.
Description
PRIOR APPLICATION
[0001] This is a continuation of U.S. patent application Ser. No.
16/655,635, entitled "Method and Apparatus for Non-exact Matching
of Addresses", filed by Richard Baker and Kaiyu Pan on Oct. 17,
2019, now U.S. Pat. No. 11,269,841, issued on Mar. 8, 2022, this
patent incorporated herein in its entirety.
BACKGROUND
Technical Field
[0002] The present disclosure relates generally to non-exact
pattern matching and more particularly to the non-exact pattern
matching of addresses and locations.
Description of the Related Art
[0003] Fraud and security are top priorities for organizations of
all sizes globally. Losses due to fraud in the banking industry
rose to $2.2 billion in 2016, up 16 percent from 2014, according to
the latest American Bankers Association (ABA) Deposit Account Fraud
Survey Report. With over 70% of companies reported to have suffered
at least one type of fraud in the past year, today's organizations
need all the protection they can get. The impact of new regulations
such as the SWIFT Customer Security Programme, Sarbanes Oxley and
Dodd-Frank, as well as increased responsibility to ensure that
financial transactions not in violation of imposed sanctions, are
of critical risk concern to enterprises of all sizes. In addition,
bank failures and the fragility of the macro-economic environment
all represent risk.
[0004] In order to meet compliance and customer expectations in the
detection of fraud, several organizations maintain lists of
entities and persons known to be associated with fraud. Other
organizations maintain lists of individuals and organizations that
are politically connected and possibly subject to bribery and
corruption.
[0005] Business applications often want to search through these
lists, which comprise a large number of `Bad Entities` to see if a
given `Target Entity` is a suspect. In this document the data store
of `Bad Entities` is considered to be the `Source` being searched.
The `Source` may be obtained from: [0006] A governmental body,
which has identified a number `Entities` that were sanctioned.
[0007] A list of politically exposed persons (PEP). [0008] A list
of bad entities identified internally by an organization or
business. [0009] An aggregation of many `Bad Guy Lists`.
[0010] Sometimes more than one `Source` may need to be checked,
because a `Bad Entity` may appear on one list but not another list.
These lists may include up to 30 million names and addresses.
[0011] While a simple search may work, it is not complete, because
small errors will prevent matches. As a result, a Levenshtein
Distance comparison and other fuzzy query techniques are performed
to match non-exact data. See U.S. Pat. No. 10,235,356, "Dual
authentication method for identifying non-exactly matching text"
for further information on non-exactly matching algorithms, said
patent incorporated herein by reference. See also U.S. patent
application Ser. No. 16/455,811, "Two Step Algorithm for Non-exact
Matching of Large Datasets".
[0012] Names alone create a complex problem to efficiently solve,
but when addresses are factored in as well, the problem becomes
even more difficult. With addresses, there are two issues. One is a
typographical error in the address text, the other is that some
locations have multiple addresses. For example, the headquarters
for the Bottomline Technologies is on the corner of Grafton Road
and Corporate Drive. The address of 255 Grafton Road in Portsmouth
NH points to the same building as 325 Corporate Drive in Portsmouth
NH. Either address could be used for Bottomline's headquarters, but
traditional address comparison software will not identify that both
addresses are the same. Some buildings are located on geographical
borders, so the town and perhaps the state could be different for
the same location. For instance, the Pheasant Lane Mall in Nashua,
N.H. was built on the New Hampshire and Massachusetts border, with
addresses in Nashua, N.H. and Tyngsborough, Mass.
[0013] Similarly, some portions of the population frequently move
around within a neighborhood, and general location may be more
helpful in identifying a match than a text address. Consider
college students, for example, who may move every semester to
another location on campus. Address names are not helpful in the
match, but the location would be helpful.
[0014] The sheer size of the lists to compare, over 30 million
names, makes this comparison computationally overwhelming,
particularly when addresses are used for the match. Cache
techniques, huge memory stores, hashing, binary search technique
fail to perform searches within a reasonable timeframe. A new
approach to comparing a name and address to a potential fraud list
is needed in the industry. The present inventions resolve this
issue.
SUMMARY OF THE INVENTIONS
[0015] A method for comparing a target address with a base address
is described herein. The method is made up of the steps of (1)
receiving the base address in text format, (2) looking up a base
location information for the base address, (3) calculating a base
geohash from the base location information, (4) receiving the
target address in text format, (5) looking up a target location
information for the target address, (6) calculating an target
geohash from the target location information, (7) subtracting the
target geohash from the base geohash, resulting in a difference,
(8) normalize the difference, and (9) comparing the difference to a
first threshold value. If the difference is less than or equal to
the first threshold value, returning an indication of a match. If
the difference is greater than the first threshold value,
calculating a score Levenshtein distance between the target address
and the base address in the target, and returning the indication of
the match if the score Levenshtein distance is greater than a
second threshold value.
[0016] In the method, the base location information, and the target
location information, each could be a longitude value and a
latitude value. The method could also include the step of
calculating the geohash by combining bits of the longitude value
with the bits of the latitude value, where even bits in the geohash
are taken for the longitude value and odd bits in the geohash are
taken for the latitude value. The method could also include the
step of taking an absolute value of the difference. The normalizing
of the difference could involve finding a most significant bit of
the difference where the most significant bit is found with a log2
function or a shift loop algorithm. The normalizing of the
difference could involve dividing the most significant bit location
value in the difference by five.
[0017] A method for locating a target address in a watch list is
also described here. The method is made up of the steps of (1)
receiving the target address in text format, (2) looking up a
target location information for the target address, (3) calculating
an target geohash from the target location information, (4) adding
a first threshold value to the target geohash resulting in a
geohash high range, (5) subtracting the first threshold value to
the target geohash resulting in a geohash low range, (6) extracting
all extracted records in the watch list, where the extracted
records contain a watch list geohash value between the geohash low
range and the geohash high range, and (7) identifying the extracted
records as matching the address target.
[0018] A method for locating a target address in a watch list could
also include the step of (8) looping through the watch list
records, if the there are no extracted records, and calculating a
score Levenshtein distance between the target address and a watch
list address in the watch list, and returning each watch list
record where the score Levenshtein distance is greater than a
second threshold value. The target information could be a longitude
value and a latitude value.
[0019] In some embodiments of the method the extracting of the
extracted records in the watch list could involve a linear search.
In another embodiment of the method the extracting of the extracted
records in the watch list includes using the geohash value as a key
to locate the watch list records. The watch list could be sorted by
the watch list geohash value.
[0020] An apparatus for determining if a target address is found in
a watch list of addresses is also described herein. The apparatus
includes a computing device with a user interface for accepting the
target address in text format. The computing device is configured
to convert the target address into target location information. The
apparatus also includes an internet, connected to the computing
device, a plurality of high performance storage devices containing
the watch list of addresses; and a high performance computing
server, connected to the internet and to the storage devices. The
high performance computing server operates software instructions to
loop through the addresses in the watch list, and for each watch
list address calculates a difference between the target address and
each watch list address, comparing the difference to a first
threshold, and indicating a match if the difference is less than or
equal to the first threshold. If the difference is greater than the
first threshold, the high performance computing server calculates a
score Levenshtein distance between the target address and each
watch list address, and indicates the match if the score
Levenshtein distance is greater than a second threshold value.
Essentially, geohash difference is used first to try to find a
match, and if not found, then the Levenshtein distance is used to
see if there is a typographical error in the addresses.
[0021] In some embodiments, the score Levenshtein distance is
calculated before the difference is calculated. In some
embodiments, the difference is calculated by subtracting a target
geohash for the target address from a base geohash for each watch
list address, and normalizing the difference. The normalizing of
the difference could involve finding a most significant bit of the
difference.
BRIEF DESCRIPTION OF THE DRAWINGS
[0022] The annexed drawings, which are not necessarily to scale,
show various aspects of the inventions in which similar reference
numerals are used to indicate the same or similar parts in the
various views.
[0023] FIG. 1 is a schematic diagram of the overall architecture of
the two step watch list search.
[0024] FIG. 2 is a flow chart of one embodiment of the search
step.
[0025] FIG. 3 is a flow chart of one embodiment of the watch list
creation.
[0026] FIG. 4 is a flow chart of one embodiment of the scoring
step.
[0027] FIG. 5 is a flow chart of one embodiment of the scoring of a
single record.
[0028] FIG. 6 is a block diagram depicting one possible hardware
embodiment of the algorithms in the proceeding figures.
[0029] FIG. 7 is a flow chart of process for adding addresses to
the watch list.
[0030] FIG. 8 is a flow chart of the search process using
addresses.
DETAILED DESCRIPTION
[0031] The present disclosure is now described in detail with
reference to the drawings. In the drawings, each element with a
reference number is similar to other elements with the same
reference number independent of any letter designation following
the reference number. In the text, a reference number with a
specific letter designation following the reference number refers
to the specific element with the number and letter designation and
a reference number without a specific letter designation refers to
all elements with the same reference number independent of any
letter designation following the reference number in the
drawings.
[0032] The present disclosure provides in one embodiment a
computer-implemented method for providing medium two step algorithm
for the non-exact pattern matching of large datasets. The method is
performed at least in part by circuitry and the data comprises a
plurality of data items.
[0033] The `Watch List Screening` (WLS) module uses a two-step
process to wade through a large number of these `Source Entities`
101,102,103,104 to find possible matches for `Target Entities` 107
being checked. The first is the `Search Step` 108 and the second is
the `Scoring Step` 110.
[0034] The first step (the `Search Step`) 108 utilizes various
searching techniques provided by the Elasticsearch (ES) platform.
One of these is a fuzzy query that uses similarity based on the
Levenshtein Edit Distance.
[0035] By carefully preparing the `Source Data` 105 to be searched,
and then indexing 106 that `Source Data` during the loading process
(using configuration options that are specified in the `ES Index
Template`) Elasticsearch can perform a search 108 and willow down
or reduce the number of possible candidates which should be scored
109. This subset list of possible candidates 109 is then passed to
the second step (the `Scoring Step`) 110.
[0036] The goal of the first step 108 is always to reduce the
number of possible candidates to be as small as possible.
[0037] Reducing the number of possible candidates is always a
balancing act. If too many `possible candidates` are gathered (too
many false positives), then the `Scoring Mechanism` 110 will be
flooded which will significantly prolong the time needed to return
a score.
[0038] Conversely, tightening the net too much during the first
step may result in missing a few possible hits in the subset list
109.
[0039] Fortunately, Elasticsearch (ES) provides a number of ways to
configure the fuzzy query used in the first step 108 to strike a
proper balance. This is explained in the Query Template--Step
1--`Search` Configuration (ES) section of this document.
[0040] The second step (the `Scoring Step`) 110 uses a `Scoring
Plug-In` 110 to score each of the possible candidates 109 returned
by the query in the first step (the `Search Step`) 108. It further
reduces the subset list of possible candidates 109 by only
returning `Suspected Entities` 111 based on close matches having
scores that are above a specified threshold. Exactly how the
`Scoring Step` 110 works will be discussed later in this
document.
[0041] Both the first step (the `Search Step`) 108 and the second
step (the `Scoring Step`) 110 are defined and configured in the
query template. The query template consists of two sections; the
first for configuring the parameters which the `Searching Tools`
use in the `Search Step` 108, and the second for configuring the
Scoring Plug-In used in the `Scoring Step` 110. In some
embodiments, the query parameters are submitted with a personal
computer or similar 601, and the query parameters, once entered,
are sent to the server 603 running the matching algorithm 100.
[0042] The first step (the `Search Step`) 108 in particular depends
on having a properly configured query template which guides 108 in
how to search against the various indexed fields for a given source
105 that are specified in the index template 106. Correspondingly,
for the first step 108 to efficiently find possible candidates 109,
it is critical to have a properly configured index template that is
tuned specifically for the `Source Data` being searched.
[0043] It is also important to properly configure the `Fields of
Interest` as part of the field map configuration for the source
being searched. This "Field Map Configuration" will be used during
the loading process to associate `Fields of Interest` in the Source
Input file 101,102,103,104, with what will be indexed using the ES
Index Template 106.
[0044] Elasticsearch is a search engine 108 that provides a
distributed, multitenant-capable full-text search engine with an
HTTP web interface and schema-free documents. It performs fuzzy
searches based on edit distance (for example, a Levenshtein
algorithm) and excels at recommending documents with very similar
structural characteristics and more narrow relatedness.
Elasticsearch performs scalable searches in near real-time.
[0045] Looking to FIG. 1, the process starts with one or more
databases of names and information regarding the people on the
various watch lists. These databases could include the OFAC List
101. This list is maintained by the US Government Department of the
Treasury, and includes individuals and companies owned or
controlled by, or acting for or on behalf of, targeted countries
(countries that US National policy or security have designated as
restricted). It also lists individuals, groups, and entities, such
as terrorists and narcotics traffickers designated under programs
that are not country-specific. Interpol, the International Criminal
Police Organization, maintains databases 102 related to the
criminal history of individuals as well as on stolen and
counterfeit documents.
[0046] There are a number of commercial sources of Politically
Exposed Persons (PEP) lists 103. A politically exposed person is
someone who is or has been entrusted with prominent public
functions by a sovereign, for example Heads of state or Heads of
government, senior politicians, senior government, judicial or
military officials, senior executives of state owned corporations,
important political party officials. Some lists include family
members or close associates, any individual publicly known, or
known by the financial institution to be a close personal or
professional associate of a politically exposed person. The PEP
lists 103 tend to be quite large, with millions or tens of millions
of records. These lists are privately maintained and sold by a
number of companies.
[0047] Internal lists 104 are lists developed by the entity used
the present inventions. This could include individuals and entities
that have committed fraud or problems for the entity, or could
include employees and officers of the entity, to prevent
self-dealing.
[0048] One of more of the OFAC list 101, the Interpol databases
102, the PEP lists 103, and the internal lists 104, perhaps also
including other databases or lists, are preprocessed and combined
into a watch list 105 and a watch index 106. FIG. 3 shows this
process. In many embodiments, the source lists 101, 102. 103, 104
are converted to comma separated lists, and each comma separated
list is preprocessed and inserted into the watch list 105 while
maintaining the watch index 106.
[0049] Often a single `Bad Entity` may be reported with a set of
related `Aliases` or AKA's (also known as). Thus, a single `Bad
Entity` may have more than one `Field Value` for a field like
`personFullname` for example. The CSV Source Input File loader,
provided with this module has the facility to load in and associate
several `Field Values` for a single `Field Name`. Other embodiments
can incorporate additional source file types, such as XML, JSON,
XLS, XLXS, PDF, .DB, .ACCDB, .NSF, .FP7 and other database file
formats.
[0050] It's important to remember that each `Bad Entity` should be
associated with a unique `Source UID` so that should a `Hit` be
scored during checking, this unique `Source UID` can be returned
and then referenced back to the source data. Normally the `Source
UID` is the first field of data in each CSV line of data being
loaded. There should be one `Source UID` per `Source Entity`. This
is expected in the system and it should be implemented this way
even when using a `Custom Field Map`.
[0051] It should be noted that even data fields which cannot be
provided on the `Target Side` may still prove valuable; they might
need to be returned (from the `Source Data`) as part of any `Watch
List Hit`.
[0052] In determining any `Custom Field Map` it must also be
ascertained which fields from the `Source Data Store` will never be
used (Don't Use field in Table 3). These are excluded in some
embodiments from the `Field Map` and also from the `CSV
Formatted--Source Input File` which will be loaded.
[0053] The more targeted the data 101, 102, 103, 104 is when
loaded, indexed, queried and then scored, the more efficient the
entire process will be.
[0054] It's also important to remember that different Source Data
Stores (Indexes) 106 can and should be used for different `Business
Applications`. Upon examining the `Source Data` it may be
determined that certain types of entities (Business vs. Person) may
need to be treated differently. Such entities may be separated into
their own CSV formatted--Source Input files 101, 102, 103, 104. For
example, if only `Sanctions Data` needs to be checked for
`Payments` then only `Sanctions Data` should be loaded into its own
dedicated CSV formatted--Source Input file.
[0055] The data input may include the first name, the middle name,
the last name, nationality, street, city, state, country, postal
code, date of birth and gender. In most implementations the data
input arrives in text format (ASCII or similar). The target
location information (latitude and longitude, or similar) may be
input 701 as well, as seen in FIG. 7, or the street, city, and
state address may be converted into a latitude and longitude 702
(or geohash or similar) using an application programming interface
(Google Maps Geocoding API--GetCoordinates(textAddress)), and that
values stored with the rest of the data 704.
[0056] Returning to FIG. 1, the watch list 105 and watch index 106
are used by the search step 108 in the search for the target 107.
FIG. 2 details the search step 108. The search step 108 searches
the watch list 105 for a subset 109 of records that could be
matches for the target 107. Once the subset list 109 is formed, the
subset list 109 is sent to the scoring step 110 for secondary
comparison to the target 107. If any of the scores are above a
threshold, then the determination 111 indicates a match. If no
scores exceed the threshold then the determination 111 indicates no
match. FIG. 4 details the Scoring Step 110.
[0057] In some embodiments, the scoring step 110 is called in
parallel with the search step 108, as a subroutine by the search
step 108 directly whenever a match is found, so that when the
subset list 109 is created, the score is stored as well. In other
embodiments, the subroutine outlined in FIG. 5, score entry 403 is
executed in parallel as called as a subroutine from the search step
108. In still another embodiment, the scoring step 110 is executed
in serial, once the search step 108 is completed.
[0058] Looking to FIG. 3, we see the algorithm for the creation of
the watch list 300. The process transforms the source lists 101,
102, 103, 104 into the watch list 105 and the watch list index 106.
The overall algorithm is a "for loop" 301,311 processing each field
of each entry in each of the source lists 101, 102, 103, 104. A
record is added to the index once the normalization is finished.
When there are no more entries to process, the routine exits and
returns the watch list 310. While the processing of all entries in
the lists 301 is not complete, the algorithm starts by removing all
capitalization 302 from each field, converting everything to lower
case. Next all repeated letters are removed 303, so, for instance,
the word "letter" becomes "leter". Then all punctuation, titles,
spaces and accents are removed 304 from each field. Articles ("a",
"an", "the") are removed. Letters with accents are converted to a
letter without the accent. For instance, a is converted to an
a.
[0059] All geographic and corporate terms are then normalized 305.
For instance, changing LLC, Corp, and similar to Inc, and changing
all references to the United States, US, USA, etc to US (this is
configurable in some embodiments). If not already available, the
address (street, city, state) is converted into location
information (base location information) such as geohash or latitude
and longitude 702, 703 using an application programming interface
(such as Google Maps Geocoding API--GetCoordinates(textAddress)).
In the preferred embodiment, the latitude and longitude is stored
with 4 decimal places, representing 11 meters (about 40 feet),
representing the size of a small lot. Existing latitude and
longitude data could also be truncated to 4 decimal places. If
using a geohash, eight characters (40 bits) are needed to get the
precision to 19 meters (about 60 feet). Additional precision in the
location data creates noise for the algorithm and consumes data
space without adding value.
[0060] Then each term or name is split into tokens 306. In some
embodiments, this split could be each word, or each syllable or
each phrase. The order of the steps in FIG. 3 can be changed
without deviating from the inventions disclosed herein.
[0061] In one embodiment, the algorithm pre-processes the terms for
Levenshtein automaton 307. This step includes creating a parametric
and generic description of states and transitions of the
Levenshtein automation. See "Fast String Correction with
Levenshtein-Automata", a paper by Klaus Schulz and Stoyan Mihov,
included herein by reference, for more information on this
step.
[0062] Finally, the algorithm builds inverted indexes 308. In doing
so, there are a few things we need to prioritize: search speed,
index compactness, and indexing speed. Changes are not important
and the time it takes for new changes is not important, because the
source lists 101, 102, 103, 104 are not updated in real time. In
some embodiments, the entire watch list 106 and watch index 106 are
regenerated to accommodate a change.
[0063] Search speed and index 106 compactness are related: when
searching over a smaller index 106, less data needs to be
processed, and more of it will fit in memory. Both, particularly
compactness, come at the cost of indexing speed, which is not
important in this system.
[0064] To minimize index 106 sizes, various compression techniques
are used. For example, when storing the postings (which can get
quite large), the algorithm does tricks like delta-encoding (e.g.,
[42, 100, 666] is stored as [42, 58, 566]), using a variable number
of bytes (so small numbers can be saved with a single byte).
[0065] Keeping the data structures small and compact means
sacrificing the possibility to efficiently update them. This
algorithm does not update the indices 106 at all: the index files
106 are immutable, i.e. they are never updated. This is quite
different to B-trees, for instance, which can be updated and often
lets you specify a fill factor to indicate how much updating you
expect.
[0066] In some embodiments, the index consists of three fields, the
term, its frequency, and its locations. Let's say we have these
three simple documents: (1) "Winter is coming.", (2) "Ours is the
fury." and (3) "The choice is yours." After some simple text
processing (lowercasing, removing punctuation and splitting words),
we can construct the "inverted index" shown in the figure.
TABLE-US-00001 TABLE 1 Term Frequency Location choice 1 3 coming 1
1 fury 1 2 is 3 1, 2, 3 ours 1 2 the 2 2, 3 winter 1 1 yours 1
3
[0067] The inverted index maps terms to locations (and possibly
positions in the locations) containing the term. Since the terms in
columns 1 and 2 (the dictionary) are sorted, we can quickly find a
term, and subsequently its occurrences in the postings-structure.
This is contrary to a "forward index", which lists terms related to
a specific location.
[0068] A simple search with multiple terms is then done by looking
up all the terms and their occurrences, and take the intersection
(for AND searches) or the union (for OR searches) of the sets of
occurrences to get the resulting list of documents. More complex
types of queries are obviously more elaborate, but the approach is
the same: first, operate on the dictionary to find candidate terms,
then on the corresponding occurrences, positions, etc.
[0069] Consequently, an index term is the unit of search. The terms
we generate dictate what types of searches we can (and cannot)
efficiently do. For example, with the dictionary in the figure
above, we can efficiently find all terms that start with a "c".
However, we cannot efficiently perform a search on everything that
contains "ours". To do so, we would have to traverse all the terms,
to find that "yours" also contains the substring. This is
prohibitively expensive when the index is not trivially small. In
terms of complexity, looking up terms by their prefix is
O(log(n))), while finding terms by an arbitrary substring is
O(n).
[0070] In other words, we can efficiently find things given term
prefixes. When all we have is an inverted index, we want everything
to look like a string prefix problem. Here are a few examples of
such transformations. Not all embodiments use all of the
examples.
[0071] 1. To find everything ending with "tastic", we can index the
reverse (e.g. "fantastic".fwdarw."citsatnaf") and search for
everything starting with "citsat".
[0072] 2. Finding substrings often involves splitting terms into
smaller terms called "n-grams". For example, "yours" can be split
into "^yo", "you", "our", "urs", "rs$", which means we would get
occurrences of "ours" by searching for "our" and "urs".
[0073] 3. For languages with compound words, like Norwegian and
German, we need to "decompound" words like "Donaudampfschiff" into
e.g. {"donau", "dampf", "schiff"} in order to find it when
searching for "schiff".
[0074] 4. Geographical coordinate points such as (42.797730,
-70.957840) can be converted into "geo hashes", in this case
"DRTTHBTJ". The longer the string, the greater the precision. The
geohash code 703 takes a binary representation of the longitude and
latitude, and combines the bits. Assuming that counting starts at 0
in the left side, the even bits are taken for the longitude code
(0111110000000), while the odd bits are taken for the latitude code
(101111001001). This operation results in the bits 01101 11111
11000 00100 00010. This value is stored and used in the present
algorithms. For display, each 5 bit grouping of the geohash is then
converted into a variant of base 32. The variant of base 32 using
all digits 0-9 and almost all lower case letters except a, i, l and
o like this:
TABLE-US-00002 TABLE 2 Decimal 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
15 Base 32 0 1 2 3 4 5 6 7 8 9 b c d e f g Decimal 16 17 18 19 20
21 22 23 24 25 26 27 28 29 30 31 Base 32 h j k m n p q r s t u v w
x y z
[0075] 5. To enable phonetic matching, which is very useful for
people's names for instance, there are algorithms like Metaphone
that convert "Smith" to {"SMO", "XMT"} and "Schmidt" to {"XMT",
"SMT"}.
[0076] 6. When dealing with numeric data (and timestamps), the
algorithm automatically generates several terms with different
precision in a trie-like fashion, so range searches can be done
efficiently. Simplified, the number 123 can be stored as
"1"-hundreds, "12"-tens and "123". Hence, searching for everything
in the range [100, 199] is therefore everything matching the
"1"-hundreds-term. This is different to searching for everything
starting with "1", of course, as that would also include "1234",
and so on.
[0077] 7. To do "Did you mean?" type searches and find spellings
that are close to the input, a "Levenshtein" automaton is built to
effectively traverse the dictionary.
[0078] Once each of the entries in the Source lists 101, 102, 103,
104 are processed, the done check 301 returns yes, and the watch
list is returned 310 from the routine.
[0079] While the above description of the watch list creation 300
is described as a series execution, in many embodiments the watch
list is created using a parallel process, where numerous processors
execute numerous processes (or threads) in parallel to create the
watch list. Other than assuring that the index is created in
coordination between the parallel processes, the remaining tasks in
the watch list creation 300 process could be executed in isolation
on separate processors and/or in separate processes.
[0080] FIG. 2 outlines the algorithm for performing the search 108
in the watch list 105 for the target 107.
[0081] The process begins by removing capitalization 201 from each
field of the target, converting all letters to lower case. The
repeat letters in the target fields are collapsed into a single
letter 202. Accents, spaces, titles and punctuation are removed
203, and geographical locations and corporation designations and
the like are converted a common terms 204. If necessary, the
longitude and latitude 802 and geohash codes 803 are calculated.
The target fields are then split into terms 205 (or syllables in
some embodiments). For example, if the target name is "Dr. Joseph
Louis Lin Manuel Hernanndez" may be converted to joseph
louislinmanuel hernandez, with the capitalization removed, the
spaces and punctuation removed from the middle names, the accent
removed from the "a", and the repeated "n" collapsed into a single
"n".
[0082] Next, the search step 108 algorithm works its way through
the entries in the watch list 105 looking for similarities. In some
embodiments, this is done with a for loop 206, 210, checking the
Levenshtein distance 207 for each entry in the watch list 105 and
scoring the record 211 using a Levenshtein distance normalized over
the length of the field in characters score. This is called the
"score Levenshtein distance" herein, and is calculated as
below:
score = 1 - Levenshtein .times. .times. ( record ) length .times.
.times. ( record ) ##EQU00001##
[0083] Each entry where the Levenshtein distance normalized score
("score Levenshtein distance") is greater that a specified
threshold 208, the entry is saved 209. In some embodiments, the
entry is scored with a weighted mean score at this point in time
209 with the algorithm shown in FIG. 5, but only for the entries
within the threshold.
[0084] In a more complicated embodiment, a hash algorithm could be
used spanning the letters on either side on the keyboard of the
target. So to search for "joseph" in the above example, the "u",
"i", "h", "k", "n" and "m" are checked with the hash, and the
following letters are checked using the complex deterministic
automation described in the "Fast String Correction with
Levenshtein-Automata", a paper by Klaus Schulz and Stoyan
Mihov.
[0085] In other embodiments, the performance of the Levenshtein
algorithms is enhanced by filtering on the lengths of the watch
list records. For instance, if the target name is 12 characters
long and the maximum Levenshtein distance parameter is set to 2,
then any watch list record that is less than 10 characters or more
than 14 characters cannot match, and are immediately eliminated
from the records that are reviewed in the loop. This technique
alone may eliminate most of the records without the computational
impact of performing a Levenshtein analysis.
[0086] Any of these approaches will result in a subset list 109 of
the watch list 105 being returned 212, where the subset list 109
has the closest matches, in some embodiments returning 100 or upto
10,000 entries out of a 30 million entry watch list 105. In many
embodiments, the loop is executed in parallel on many processors
using many processes or threads to process the watch list.
[0087] The scoring step is diagramed in FIG. 4 and FIG. 5. The
overall algorithm is shown in FIG. 4. The subset list 109 and the
target 107 are the inputs to this step. The process loops 401, 406
through each record in the subset list 109, scoring each entry 403
with the scoring algorithm described in FIG. 5. If the score is
greater than the threshold 404, then the score is stored 405 with
the record. If the score is less than or equal to the threshold
404, then no score is returned 410 (or no record is returned), and,
in some embodiments, the record is deleted from the subset list
109. After all records are processed, the records that have scores
greater than the threshold are returned 410. The threshold is a
parameter that can be adjusted based on each implementation. This
parameter determine how close a match needs to be to be considered
a match. In some embodiments, the loop is executed in parallel on
many processors using many processes or threads to process the
subset list.
[0088] The scoring process can be summarized as follows:
[0089] Step 1--The system will first check each of the required
fields 510, one by one, individually. If even one of the required
fields has a score which falls below the specified field-level
threshold, then no score will be returned. (In Table 3, all fields
in the "firstname-lastname-score"--`Query Template` have a default
field-level threshold. If no field-level default was defined then
the overall threshold would be used.)
[0090] Step 2--The system will calculate the weighted mean of the
required fields 503 using the field weights assigned and the field
scores calculated for each of the required fields.
[0091] Step 3--The system will see if there are any boost fields
with scores that are above their field-level threshold and ALSO
above the weighted mean of the required fields and use only those
scores to boost the aggregated required fields score 520.
[0092] Step 4--A new boosted aggregated score will then be
calculated by taking the weighted mean of the required field scores
and the boost field scores discovered in Step 3 526. This will
improve or "boost" the score calculated in Step 2.
[0093] It is the boosted aggregated score which will be returned
(assuming it is above the specified threshold).
[0094] FIG. 5 details the score name 403 step of FIG. 4. The score
name step 403 takes an entry from the subset list 501 and the
target 107 and scores the closeness of the match between the two.
This match is done in two stages. First that required fields each
compared 510 using a Levenshtein type score and then the boost
fields are compared 520 if they will increase the score.
[0095] The score name function 403 steps through each character in
each field, one by one, for each target 107 to be scored, and notes
each difference against the subset list entity 501 it is being
compared against. If no match could be found in the search step 108
for a given target 107 then no scoring will be done for that target
107 and the assumption will be that no hit was detected.
[0096] The comparison, character by character, between the target
107 fields and the subset list field 501 field provided in a loaded
`Bad Guy List` will be done after all of the analyzer filters
specified for that source 501 field have been applied to both
sides. For the source, the analyzer filers are run when the entry
is indexed. For target, this happens before the score is
calculated.
[0097] If every character matches perfectly, then a `Score of 1.00`
will be returned 527. If "Some but not all" characters match, then
a "Score" will be calculated to reflect how close the match is. A
higher score will indicate a better `match`. A lower score will
indicate a worse `match`.
[0098] A threshold parameter set threshold floor of what the
scoring name 403 can return. The threshold for the search step 108,
is a similar value. To make sure that only meaningful `hits` are
returned, an overall default threshold will be the same value for
the search step 108 and the scoring step 110.
[0099] Here, the threshold parameter sets the minimum score that a
target 107 must receive before it is returned in the response as a
`hit`. A weighted mean of each field score is calculated 503, 526
to obtain a single value for the overall score. If the overall
score is below the threshold value then no result is returned in
the response 516, 527. So as poor matches are detected, they can be
held back, and not flood the `Results` with too many
`False-Positives`.
[0100] The correct `Overall Threshold` to use varies widely from
organization to organization. Very risk adverse organizations,
which are afraid of missing too many `Real Hits` may set this
`Overall Minimum Threshold` as low as 80% but that could result in
a very large number of `Possible Hits` to review, and more
false-positive hits. Less risk adverse organizations, might push
this `Overall Minimum Threshold` a bit higher, and accept the added
risk of missing some `Real Hits` to minimize the number of
false-positive hits it will need to review and discard.
[0101] In some embodiments, there is a table of field related
parameters (for example, see Table 3), and in other embodiments,
the settings are hard coded into the code. Each field is designated
as either Required, Boost, or Don't Care. The Don't Care fields are
ignored, and in some embodiments are eliminated from the watch list
at the watch list creation 300 stage. These Don't Care fields are
similarly eliminated as the target 107 is entered. If the field is
a Don't Care, there is no need to provide weight or threshold. The
Required fields are processed in the required field stage 510 and
the Boost fields are processed in the boost fields stage 520. The
table also includes the weight and the threshold for each field.
These values are needed in the calculations below.
TABLE-US-00003 TABLE 3 firstname middlename lastname SSN street
city state weight 1 0.5 1 0.2 .2 0.2 0.2 threshold 0.8 0.8 0.8 0.8
.05 0.8 0.8 designation REQUIRED BOOST REQUIRED BOOST BOOST BOOST
BOOST
[0102] Looking to FIG. 5, the required fields 510 are processed
first. Starting with the first field, the first name in the example
in Table 3, the field score is calculated 511. This calculation is
one minus the Levinstein distance between the subset list entry 501
field and the target field 107 divided by the number of characters
in the target 107.
score = x = 1 - Levenshtein .times. .times. ( subset . first
.times. .times. name , target . firstname ) length .times. .times.
( target . fir .times. s .times. t .times. n .times. a .times. m
.times. e ) ##EQU00002##
[0103] If the field score is less than or equal to the threshold
512, then there is no match between this record and the target. The
score name 403 returns an indication that there is no match 516,
perhaps by returning a zero.
[0104] For all address calculations, first the geohash location is
absolute value of the difference between the subset geohash and the
target geohash is calculated. Then the most significant bit of the
difference is identified and that number converted into the score:
[0105] int msb=(int)(log2(difference)); [0106] int score=msb/5;
[0107] In some embodiments, the log2( ) function is calculated with
a shift left loop, looping until the most significant bit is set,
and using the count of the number of loops, subtracted from the
number of bits in the difference. [0108] for (int i=31; i>=0;
1--) [0109] if (difference=<<1) & 0x8000000) [0110]
return i/5;
[0111] If the score for the location is greater than the threshold,
then we could calculate the Levenshtein score for the text address
(street, city and state) as above. The best score from either the
Levenshtein or the geohash location is used for the field
score.
[0112] If the field score is greater than the threshold 512, then
store the field score in an array 513 for later processing. Then
check to see if there are any further required fields to process
514. If so, get the next field 515 and repeat the loop at the
calculation of the field score 511.
[0113] If there are not more required fields 514 to process, then
calculate the weighted mean score 503 by summing the product of
multiplying each field score by the field weight and dividing the
sum by the sum of the weights to create the weighted mean score,,
where w is the field weight and x is the field score.
x _ = W i .times. x i W i = w 1 .times. x 1 + w 2 .times. x 2 + + w
n .times. x n w 1 + w 2 + + w n ##EQU00003##
[0114] Once the weighted mean is calculated, the boost fields are
analyzed 520. The boost fields will only be used to increase the
match score. They are not used to diminish the match score found in
the required fields analysis 510. As such, they are only
incorporated in the formula if the boost score exceeds the weighted
mean 503.
[0115] The boost field analysis 520 begins by checking to see if
there are any boost fields to process 521. If so, the field score
is calculated 522. This calculation is one minus the Levinstein
distance between the subset list entry 501 field and the target
field 107 divided by the number of characters in the target
107.
score = x = 1 - Levenshtein .times. .times. ( subset . ssn , target
. ssn ) length .times. .times. ( target . ssn ) ##EQU00004##
[0116] When checking the field score 523 the score is used only if
it is greater than the threshold and greater than the weighted
mean, x. If the field score is less than or equal to the threshold
or less than or equal to the weighted mean x 523, then the field is
ignored, the next field is retrieved 525 and the loop repeats the
test to determine if there are more boost fields to review 521.
[0117] For all address calculations, first the geohash location is
absolute value of the difference between the subset geohash and the
target geohash is calculated. Then the most significant bit of the
difference is identified and that number normalized into the score
by dividing by 5: [0118] int msb=(int)(log2(difference)); [0119]
int score=msb/5;
[0120] If the score for the location is greater than the threshold,
in some embodiments then we calculate the Levenshtein score for the
text address (street, city and state) as above. The best score from
either the Levenshtein or the geohash location is used for the
field score.
[0121] If the field score is greater than the threshold and greater
that the weighted mean x 512, then the field score is stored in an
array 524 for later processing. Then, get the next field 525 and
repeat the loop at the calculation of the field score 521.
[0122] If there are not more required fields 521 to process, then
calculate the aggregated score 526 by summing the product of
multiplying each field score by the field weight and dividing the
sum by the sum of the weights to create the weighted mean score,,
where w is the field weight and x is the field score.
x _ = W i .times. x i W i = w 1 .times. x 1 + w 2 .times. x 2 + + w
n .times. x n w 1 + w 2 + + w n ##EQU00005##
[0123] Once the aggregated mean score is calculated 526, return
this value to the calling routine 527.
[0124] FIG. 6 shows one possible hardware embodiment for this
invention. A personal computer, laptop, notebook, tablet, smart
phone, cell phone, smart watch or similar 601 is used to provide a
visual input to the process. In some embodiments, the target 107 is
entered through this computing device 601. The source lists 101,
102, 103, 104 may also be selected in some embodiments. The target
and the lists are sent to a special purpose, high performance
computing server 603. In some embodiments, the high performance
computing server 603 has a high performance floating point
processor, more memory than is found on a typical computer, and
additional cache memory so that the huge source lists can be stored
(at least in part) and processed. The high performance computing
server 603 has a large number of parallel processors in some
embodiments to allow for the parallel execution of the watch list
creation 300, the search step 108, and the scoring step 110. This
high performance computing server 603 is electrically connected to
one or more high performance storage devices 604 (such as disk
drives, optical storage, solid state memory, etc.) for storing the
watch list 105, the index 106, and the subset list 109. In some
embodiments, the computing device 601 communicates with the high
performance computing server 603 over the internet 602.
[0125] In some embodiments, the high performance computing server
603 and the one or more high performance storage devices 604 could
be distributed over the internet 602, with the functions described
in FIG. 1 distributed over a plurality of servers 603 and storage
devices 604. The distribution could be along functional lines in
some embodiments, with the watch list creation 300 performed on one
server, the search step 108 performed on another server, and the
scoring step 110 performed on still a third server.
[0126] In another distributed embodiment, a plurality of servers
603 could perform all of the steps 300, 108, 110 on a portion of
the watch list 105 and/or on a portion of the subset 109.
[0127] FIG. 7 show how an address entry is added to the watch list
700. In most embodiments, the address is one of many elements of
the watch list record, but the address needs special processing.
First of all, the text address is converted into latitude or
longitude 702 using an application programming interface (such as
Google Maps Geocoding API--GetCoordinates(textAddress)). Since this
could be a time consuming processes, in many embodiments it is
performed once per address and the latitude and longitude stored
for future use with this address. In the preferred embodiment, the
latitude and longitude is stored with 4 decimal places,
representing 11 meters (about 40 feet), representing the size of a
small lot. If the address cannot be converted, perhaps dues to a
mistyped address, several options are available. In some
embodiments, the user is asked to double check and reenter the
address. In an automated process, other tools could be used to
automatically correct the address, such as using various address
auto-correction facilities available on the internet (google maps,
US Post Office tools, etc.). If none of these work, an error code
is stored for the latitude and longitude, perhaps a 0xFFFFFFFF
value.
[0128] The geohash code 703 is created by taking a binary
representation of the longitude and latitude, and combining the
bits. Assuming that counting starts at 0 in the left side, the even
bits are taken for the longitude code (0111110000000), while the
odd bits are taken for the latitude code (101111001001). This
operation results in the bits 01101 11111 11000 00100 00010. This
value is stored 704 and returned 705.
[0129] FIG. 8 explains the lookup of the address in the watch list
800. In some embodiments, this is run before or as part of the
search step 108. The target 107 is sent to the function as a text
target address 801. First of all, the target text address is
converted into latitude or longitude 802 using an application
programming interface (such as Google Maps Geocoding
API--GetCoordinates(textAddress)). In the preferred embodiment, the
latitude and longitude is stored with 4 decimal places,
representing 11 meters (about 40 feet), representing the size of a
small lot. If the address cannot be converted, perhaps dues to a
mistyped address, several options are available. In some
embodiments, the user is asked to double check and reenter the
target address. In an automated process, other tools could be used
to automatically correct the address, such as using various address
auto-correction facilities available on the internet (google maps,
US Post Office tools, etc). If none of these work 807, then no
records are searched and the algorithm proceeds directly to the
Levenshtein search step 108.
[0130] If there are no errors in the conversion 807, the geohash
code 803 is created by taking a binary representation of the
longitude and latitude, and combining the bits. Assuming that
counting starts at 0 in the left side, the even bits are taken for
the longitude code (0111110000000), while the odd bits are taken
for the latitude code (101111001001). This operation results in the
bits 01101 11111 11000 00100 00010.
[0131] Next, the geohash range is created 804 by adding and
subtracting a threshold to the geohash, creating a geohash low
range and a geohash high range value. Using the geohash low range
and geohash high range values, all of the watch list records within
the geohash range are extracted 808 (extracted records) using the
watch list addresses as converted into geohashes. In one
embodiment, the watch list has been stored as a sorted list, sorted
by geohash values. In other embodiments, the geohash is a database
key or as a lookup table. In these embodiments, the records are
pulled by indexing into the list and quickly identifying the
records with geohashes within the range. In other embodiments, a
binary search is performed on the sorted geohash address list. In
other embodiments, a linear search is used to find watch list
geohashes within the range. In still other embodiments, a tree type
search could be used, with the geohash divided by a constant to
derive a set of geohash address buckets that are then linearly
searched.
[0132] If records are found 805, they are returned for further
processing 806, perhaps checking other information in the record
for matches, or using this list as the subset 109 for processing by
the scoring step 110.
[0133] If no records are found 805, then the process proceeds with
the Levenshtein search step 108, and this result is returned
806.
[0134] It should be appreciated that many of the elements discussed
in this specification may be implemented in a hardware circuit(s),
a circuitry executing software code or instructions which are
encoded within computer readable media accessible to the circuitry,
or a combination of a hardware circuit(s) and a circuitry or
control block of an integrated circuit executing machine readable
code encoded within a computer readable media. As such, the term
circuit, module, server, application, or other equivalent
description of an element as used throughout this specification is,
unless otherwise indicated, intended to encompass a hardware
circuit (whether discrete elements or an integrated circuit block),
a circuitry or control block executing code encoded in a computer
readable media, or a combination of a hardware circuit(s) and a
circuitry and/or control block executing such code.
[0135] All ranges and ratio limits disclosed in the specification
and claims may be combined in any manner. Unless specifically
stated otherwise, references to "a," "an," and/or "the" may include
one or more than one, and that reference to an item in the singular
may also include the item in the plural.
[0136] Although the inventions have been shown and described with
respect to a certain embodiment or embodiments, equivalent
alterations and modifications will occur to others skilled in the
art upon the reading and understanding of this specification and
the annexed drawings. In particular regard to the various functions
performed by the above described elements (components, assemblies,
devices, compositions, etc.), the terms (including a reference to a
"means") used to describe such elements are intended to correspond,
unless otherwise indicated, to any element which performs the
specified function of the described element (i.e., that is
functionally equivalent), even though not structurally equivalent
to the disclosed structure which performs the function in the
herein illustrated exemplary embodiment or embodiments of the
inventions. In addition, while a particular feature of the
inventions may have been described above with respect to only one
or more of several illustrated embodiments, such feature may be
combined with one or more other features of the other embodiments,
as may be desired and advantageous for any given or particular
application.
* * * * *