U.S. patent application number 17/171670 was filed with the patent office on 2021-08-05 for methods, systems and apparatus to improve multi-demographic modeling efficiency.
The applicant listed for this patent is The Nielsen Company (US), LLC. Invention is credited to Xiaoqi Cui, Ludo Daemen, Michael Sheppard.
Application Number | 20210241157 17/171670 |
Document ID | / |
Family ID | 1000005524932 |
Filed Date | 2021-08-05 |
United States Patent
Application |
20210241157 |
Kind Code |
A1 |
Sheppard; Michael ; et
al. |
August 5, 2021 |
METHODS, SYSTEMS AND APPARATUS TO IMPROVE MULTI-DEMOGRAPHIC
MODELING EFFICIENCY
Abstract
Methods, apparatus, systems and articles of manufacture are
disclosed to improve multi-demographic modeling efficiency. An
example apparatus includes a feature set aggregator to segregate
training data based on feature sets of interest, and to identify
households that participate in at least one of the feature sets of
interest, a class enumerator to reduce multi-demographic model
iterations by enumerating demographic combinations for the
identified households, the enumerated demographic combinations
including a single identifier to represent a combination of two or
more demographic categories, and a modeling engine to generate
training coefficients associated with respective ones of the
enumerated demographic combinations.
Inventors: |
Sheppard; Michael; (Holland,
MI) ; Daemen; Ludo; (Duffel, BE) ; Cui;
Xiaoqi; (Darien, IL) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
The Nielsen Company (US), LLC |
New York |
NY |
US |
|
|
Family ID: |
1000005524932 |
Appl. No.: |
17/171670 |
Filed: |
February 9, 2021 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
15359971 |
Nov 23, 2016 |
10943175 |
|
|
17171670 |
|
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06N 7/005 20130101;
G06Q 30/0201 20130101 |
International
Class: |
G06N 7/00 20060101
G06N007/00; G06Q 30/02 20060101 G06Q030/02 |
Claims
1. (canceled)
2.-20. (canceled)
21. An apparatus to reduce computational resources for a
multi-demographic training model, comprising: memory; and at least
one processor to execute computer readable instructions to at
least: reduce multi-demographic model iterations by enumerating
demographic combinations for households that access at least one of
a plurality of features represented in training data, the
enumerated demographic combinations including an identifier, the
identifier to represent a combination of two or more demographic
categories and to include cartesian representations corresponding
to occurrences or non-occurrences of a demographic category of
interest, the demographic category of interest to be selected from
the two or more demographic categories; and generate probability
values for ones of the enumerated demographic combinations based on
training coefficients, the probability values to represent
likelihoods that the ones of the enumerated demographic
combinations accessed the at least one of the plurality of
features, the training coefficients corresponding to the ones of
the enumerated demographic combinations, the ones of the enumerated
demographic combinations associated with (a) a training model and
(b) a third-party model.
22. The apparatus as defined in claim 21, wherein the at least one
processor is to execute the computer readable instructions to
identify ones of the households that access the at least one of the
features of interest.
23. The apparatus as defined in claim 22, wherein the at least one
processor is to execute the computer readable instructions to
associate demographic combinations with the ones of the
households.
24. The apparatus as defined in claim 21, wherein the at least one
processor is to execute the computer readable instructions to
generate the training coefficients based on a multinomial logistic
regression with: (a) the training data, and (b) the ones of the
enumerated demographic combinations.
25. The apparatus as defined in claim 21, wherein the at least one
processor is to execute the computer readable instructions to
revert the probability values for respective ones of the enumerated
demographic combinations to second probability values associated
with individual demographic components.
26. The apparatus as defined in claim 21, wherein the at least one
processor is to execute the computer readable instructions to
determine probability value differences between the training model
and the third-party model.
27. The apparatus as defined in claim 26, wherein the at least one
processor is to execute the computer readable instructions to
determine a trust metric for the third-party model based on a
comparison of the probability value differences to a threshold, the
trust metric to indicate the third-party model is reliable.
28. A tangible computer-readable storage medium comprising
instructions that, when executed, cause a processor to, at least:
reduce multi-demographic model iterations by enumerating
demographic combinations for households that access at least one of
a plurality of features represented in training data, the
enumerated demographic combinations including an identifier, the
identifier to represent a combination of two or more demographic
categories and to include cartesian representations corresponding
to occurrences or non-occurrences of a demographic category of
interest, the demographic category of interest to be selected from
the two or more demographic categories; and generate probability
values for ones of the enumerated demographic combinations based on
training coefficients, the probability values to represent
likelihoods that the ones of the enumerated demographic
combinations accessed the at least one of the plurality of
features, the training coefficients corresponding to the ones of
the enumerated demographic combinations, the ones of the enumerated
demographic combinations associated with (a) a training model and
(b) a third-party model.
29. The computer-readable medium as defined in claim 28, wherein
the instructions are to cause the processor to identify ones of the
households that access the at least one of the features of
interest.
30. The computer-readable medium as defined in claim 29, wherein
the instructions are to cause the processor to associate
demographic combinations with the ones of the households.
31. The computer-readable medium as defined in claim 28, wherein
the instructions are to cause the processor to generate the
training coefficients based on a multinomial logistic regression
with: (a) the training data, and (b) the ones of the enumerated
demographic combinations.
32. The computer-readable medium as defined in claim 28, wherein
the instructions are to cause the processor to revert the
probability values for respective ones of the enumerated
demographic combinations to second probability values associated
with individual demographic components.
33. The computer-readable medium as defined in claim 28, wherein
the instructions are to cause the processor to determine
probability value differences between the training model and the
third-party model.
34. The computer-readable medium as defined in claim 33, wherein
the instructions are to cause the processor to determine a trust
metric for the third-party model based on a comparison of the
probability value differences to a threshold, the trust metric to
indicate the third-party model is reliable.
35. An apparatus to reduce computational resources for a
multi-demographic training model, comprising: means for reducing
multi-demographic model iterations by enumerating demographic
combinations for households that access at least one of a plurality
of features represented in training data, the enumerated
demographic combinations including an identifier, the identifier to
represent a combination of two or more demographic categories and
to include cartesian representations corresponding to occurrences
or non-occurrences of a demographic category of interest, the
demographic category of interest to be selected from the two or
more demographic categories; and means for modeling to generate
probability values for ones of the enumerated demographic
combinations based on training coefficients, the probability values
to represent likelihoods that the ones of the enumerated
demographic combinations accessed the at least one of the plurality
of features, the training coefficients corresponding to the ones of
the enumerated demographic combinations, the ones of the enumerated
demographic combinations associated with (a) a training model and
(b) a third-party model.
36. The apparatus as defined in claim 35, further including means
for aggregating to identify ones of the households that access the
at least one of the features of interest.
37. The apparatus as defined in claim 36, wherein the modeling
means is to generate the training coefficients based on a
multinomial logistic regression with: (a) the training data, and
(b) the ones of the enumerated demographic combinations.
38. The apparatus as defined in claim 35, further including means
for reverting the probability values for respective ones of the
enumerated demographic combinations to second probability values
associated with individual demographic components.
39. The apparatus as defined in claim 38, wherein the reverting
means is to determine probability value differences between the
training model and the third-party model.
40. The apparatus as defined in claim 39, wherein the reverting
means is to determine a trust metric for the third-party model
based on a comparison of the probability value differences to a
threshold, the trust metric to indicate the third-party model is
reliable.
Description
FIELD OF THE DISCLOSURE
[0001] This disclosure relates generally to demographic modeling,
and, more particularly, to methods, systems and apparatus to
improve multi-demographic modeling efficiency.
BACKGROUND
[0002] In recent years, marketing analysts have applied modeling
techniques to determine household demographic compositions. Such
modeling techniques operate on one or more types of available data
sources, such as survey data and panelist data, the latter of which
includes a degree of measurement accuracy that can be used for
predictions and/or projections that meet expectations of
statistical significance.
BRIEF DESCRIPTION OF THE DRAWINGS
[0003] FIG. 1 is a schematic illustration of a demographic modeling
system constructed in accordance with the teachings of this
disclosure to improve multi-demographic modeling efficiency.
[0004] FIG. 2 is an example training data set analyzed by the
example demographic modeling system of FIG. 1 to estimate household
demographic information with binary demographic indicators.
[0005] FIG. 3 is an example training data set analyzed by the
example demographic modeling system of FIG. 1 to estimate household
demographic information with household member count demographic
indicators.
[0006] FIG. 4 is an example training data set analyzed by the
example demographic modeling system of FIG. 1 to estimate household
demographic information in connection with third party models.
[0007] FIGS. 5-8 are flowcharts representative of example machine
readable instructions that may be executed to implement the example
demographic modeling system and/or the example demographic modeling
manager of FIG. 1 to improve multi-demographic modeling
efficiency.
[0008] FIG. 9 is a block diagram of an example processor platform
structured to execute the example machine readable instructions of
FIGS. 5-8 to implement the example demographic modeling system of
FIG. 1.
DETAILED DESCRIPTION
[0009] Marketing analysts may estimate household demographic
compositions based on viewing data. For example, by applying
retrieved and/or otherwise receiving viewing data (e.g., from
panelist households), models may be applied to determine the
presence or absence of a particular demographic of interest within
households of interest. While such modeling efforts reveal valuable
demographic insight for the demographic of interest for a
particular household, the analyst must then re-run another model
for other demographics of interest in a demographic-by-demographic
manner. In some examples, marketing analysists seek household
demographic information for many different demographic categories
of interest that are typically found in households. For instance,
some households include two parents and one or more children, which
requires modeling analysis for (a) the female (mother) of the
household for any number of age categories (e.g., 19-23, 24-27,
28-31, etc.), the male (father) of the household for any number of
(similar) age categories, and (c) the one or more children for any
number of age categories (e.g., 0-6 months, 7 months-1 year, 2-5,
6-9, etc.). For each household member, a separate model is
typically executed to determine demographic details for each
candidate age category of interest. Accordingly, for even a
relatively small number of household members, multiple modeling
iterations are required to obtain demographic estimations in view
of the viewing data.
[0010] Additionally, after any number of modeling iterations occur
for each candidate member of a household to determine individual
demographics of that household, the analyst must also perform
additional modeling efforts to identify potential correlation
effects. As such, modeling processing requirements dramatically
increase as a number of household members increases.
[0011] Methods, systems, apparatus and articles of manufacture
disclosed herein improve multi-demographic modeling efficiency.
Unlike traditional approaches to estimate demographics with viewer
data with model iterations for (a) each demographic of interest and
(b) combining the model iterations with collinearity modeling
iteration(s) and/or correction efforts, examples disclosed herein
apply a single truth model to the viewer data to generate truth
coefficients for classes of demographic combinations. Examples
disclosed herein model the classes one time to allow test data to
be applied to the model to determine corresponding demographic
probabilities for demographics of interest without requiring
individualized models to be calculated and/or otherwise applied for
each demographic of interest. As such, examples disclosed herein
facilitate technical benefits including, but not limited to,
reducing a number of computational models applied during marketing
research, and reducing an amount of computational resource
utilization during marketing research efforts. By reducing an
application of numerous models and/or modeling iterations on a
demographic-by-demographic manner, examples disclosed herein
improve computational efficiency (e.g., reduce or minimize needed
computational resources) as well as improve marketing analyst
efficiency (e.g., reduce or minimize an amount of time preparing a
marketing model) when performing marketing research.
[0012] FIG. 1 is a schematic illustration of a demographic modeling
system 100. In the illustrated example of FIG. 1, the demographic
modeling system 100 includes a demographic modeling manager 102
communicatively connected to one or more network(s) 104 that is/are
communicatively connected to a truth data store 106 and a test data
store 108. The example truth data store 106 includes panelist data
of viewing behavior in which a detailed household composition is
known. In some examples, household viewing behaviors stored in the
example truth data store 106 include tuning behavior data (e.g.,
channel tuned, channel change data, duration (e.g., in minutes) of
channel viewing, etc.) associated with set-top-boxes (STBs).
Additionally, the example tuning behavior data is associated with
household demographics (e.g., number of household members,
household member age information, household member gender
information, household member income information, etc.). In some
examples, the tuning behavior is derived from panelist households
with a PeopleMeter.RTM. (PM) metering device installed by an
audience measurement entity (e.g., The Nielsen Company (US), LLC,
Experian, etc.) that monitors media exposure data (e.g., viewing
data, listening data, gaming activity, etc.) of an enlisted
audience. Audience measurement typically involves determining the
identity of the media being displayed and/or otherwise consumed on
a media presentation device (e.g., a television), determining
demographic information of an audience, and/or determining which
members of a household are to be associated with exposure to
particular media. Some audience measurement systems physically
connect to the media presentation device, such as the television,
to identify which channel is being tuned or which program is tuned
by capturing a channel number, audio signatures and/or codes
identifying (directly or indirectly) the programming being
displayed by the media presentation device. Additionally, some
audience measurement systems prompt and/or accept audience member
input to reveal which household member is currently exposed to the
media presented by the media presentation device.
[0013] The example test data store 108 includes audience viewing
behavior information from one or more other sources, some of which
may not be managed to a statistical rigor expected by panelist data
sources. In some examples, the audience viewing behavior
information stored in the test data store 108 is devoid of
associated demographic information, but does include tuning
behavior information. In still other examples, the audience viewing
behavior information stored in the test data store 108 includes
third party models of demographic viewing behavior. Some such third
party models may be developed by audience measurement entities
and/or personnel unassociated with the market analyst using the
example demographic modeling system 100 of FIG. 1. However, some
such third party models may have particular strengths and/or
capabilities for accurately predicting, projecting and/or otherwise
estimating viewing behaviors for particular demographics. As
described in further detail below, example third party models
stored in the example test data store 108 may be applied to
examples disclosed herein to derive model corrections of one or
more models generated by the example demographic modeling manager
102.
[0014] The example demographic modeling manager 102 of FIG. 1 also
includes an example data retriever 110, an example feature set
segregator 112, an example household data associator 114, an
example class enumerator 116, an example Cartesian combiner 118, an
example modeling engine 120, an example linear combiner 122 and an
example condition engine 124. In operation, the example data
retriever 110 retrieves training data from the example truth data
store 106. The training data may be associated with viewership data
and associated demographic information from a geography of
interest. In some examples, the marketing analyst is interested in
understanding demographic information associated with one or more
particular features of the viewing audience, such as one or more
particular programs that have been tuned by and/or otherwise
consumed by audience members. Example features include, but are not
limited to particular programs of interest (e.g., broadcast news
programs, situational comedies), particular movies of interest,
particular songs, etc. Depending on the features of interest
selected by the marketing analyst, the example feature set
segregator 112 segregates the retrieved training data, and
identifies households that include such activity, as shown in the
illustrated example of FIG. 2.
[0015] FIG. 2 illustrates example training data 200 retrieved by
the example data retriever 110 (e.g., from the example truth data
store 106) and segregated by five (5) example features of interest
202. The example five features of interest 202 may be indicative of
five television shows for which tuning data occurred for
twenty-five (25) example households 204. For each intersecting
household row 204 and feature column 202, an example value of
quarter hours of viewing is shown. For instance, in the illustrated
example of FIG. 2, a first household 206 viewed ten (10) quarter
hours of feature 1, four (4) quarter hours of feature 2, eight (8)
quarter hours of feature 3, five (5) quarter hours of feature 4,
and five (5) quarter hours of feature 5.
[0016] The example household data associator 114 associates
available demographic data 208 with the respective households 204.
In the illustrated example of FIG. 2, two demographics are shown as
"D1" 210 and "D2" 212. While the illustrated example of FIG. 2
includes 25 households, five features and two demographics of
interest, such examples are included for illustration and not
limitation. Example disclosed herein may include any number of
households, features and/or demographic of interest to improve
multi-demographic modeling efficiency. The example demographic data
208 includes values of one (1) or zero (0) to indicate whether the
corresponding household included a particular demographic of
interest when the feature viewing behavior occurred. Demographic
indicators of one (1) reflect the presence of that particular
demographic, while demographic indicators of zero (0) indicate no
viewing behavior of that particular demographic. For example, the
first household 206 includes a value of 1 for demographic D1 and D2
to indicate that household members of these demographics
participated in the corresponding viewing behaviors, and a second
household 213 includes a value of 0 for demographic D1 to indicate
that no household member fitting the qualifications for D1
participated in the corresponding viewing behaviors, while
demographic D2 did participate in those viewing behaviors.
[0017] As described above, traditional approaches to developing a
training model included the application of training data on each
demographic of interest individually to generate model coefficients
that might be used for projection purposes. When one demographic of
interest was modeled to generate corresponding coefficient values,
the marketing analyst proceeded to identify the next demographic of
interest, apply a corresponding model, and generate corresponding
coefficients for that other demographic of interest. After all
demographics of interest were modeled and corresponding
coefficients were generated, the analyst again applied such
training data and coefficients to another model to consider
potential effects of collinearity. Accordingly, such traditional
approaches included multiple iterations of individualized modeling
efforts that each consume processing resources, and analyst
micro-management.
[0018] Unlike traditional approaches to estimate household
demographics to develop a model for predictions and/or projections,
examples disclosed herein enumerate unique combinations of
participating household demographics in connection with the
training data 200. The example class enumerator 116 enumerates each
unique demographic combination to a class representation. In the
illustrated example of FIG. 2, because there are two separate
demographics of interest associated with the training data 200,
there are three (3) possible unique combinations. In particular,
the example class enumerator 116 creates a first class (Class 1) to
identify circumstances in which D1 does not participate in the
viewing behavior, but D2 does participate in the viewing behavior.
Additionally, the example class enumerator 116 creates a second
class (Class 2) to identify circumstances in which D1 participates
in the viewing behavior, but D2 does not participate in the viewing
behavior. Finally, the example class enumerator 116 creates a third
class (Class 3) to identify circumstances in which both D1 and D2
participate in the viewing behavior. Note that examples disclosed
herein do not consider circumstances in which neither D1 nor D2
participate in the viewing behavior, as such circumstances do not
contribute to the modeling effort.
[0019] The example Cartesian combiner 118 assigns Cartesian
products of candidate combinations 214, in which coordinates are in
a format (D1,D2) having a binary indicator of one (1) or zero (0)
to indicate the presence or absence, respectively, of household
members associated with demographic D1 and/or D2. In the
illustrated example of FIG. 2, the Cartesian combiner 118 assigns
(0,1) as Class 1 to reflect circumstances where D1 does not
participate in the viewing behavior, but D2 does participate in the
viewing behavior. Additionally, the Cartesian combiner 118 assigns
(1,0) as Class 2 to reflect circumstances where D1 does participate
in the viewing behavior, but D2 does not participate in the viewing
behavior. Finally, the Cartesian combiner 118 assigns (1,1) as
Class 3 to reflect circumstances where both D1 and D2 participate
in the viewing behavior.
[0020] The example modeling engine 120 performs a logistic
regression of the enumerated class representations to generate
probability splits for each class representation. As used herein,
application of the logistic regression includes a multinomial
regression, which is sometimes referred to herein as a multinomial
logistic regression. As shown in the illustrated example of FIG. 2,
for each intersecting Cartesian product of candidate combinations
214 and household 204, the example modeling engine 120 generates
probability values (e.g., Class 1 probability values, Class 2
probability values and Class 3 probability values). The example
first household 206 includes a zero percent probability (see item
216) that the Class 1 combination (that is, no participation by a
household member matching demographic D1 and participation by a
household member matching demographic D2) occurs, a 29% probability
(see item 218) that the Class 2 combination occurs, and a 71%
probability (see item 220) that the Class 3 combination occurs.
When the example modeling engine 120 applies the logistic
regression to the example training data 200 in connection with the
enumerated classes 214, corresponding coefficients for each
enumerated class of interest result. In the illustrated example of
FIG. 2, the modeling engine 120 calculates a Class 1 model
coefficient "A" (see item 222), a Class 2 model coefficient "B"
(see item 224), and a Class 3 model coefficient "C" (see item 226).
Such corresponding coefficients are later used in connection with
test data and/or test sets for the purpose of demographic
projections with data for which the underlying demographic data is
unknown or not otherwise verified as true, as described in further
detail below.
[0021] To estimate and/or otherwise predict a probability of a
demographic of interest participating in a household of interest,
the example linear combiner 122 applies linear combinations of the
example corresponding class probabilities (e.g., for the example
first household 206 the Class 1 probability of 0% (see item 216),
the Class 2 probability of 29% (see item 218), and the Class 3
probability of 71% (see item 220) are used). To illustrate, the
example linear combiner 122 selects a household of interest (e.g.,
the first household 206) and a demographic of interest for which a
predicted probability value is desired (e.g., demographic D1).
Based on the selected demographic of interest, the example linear
combiner 122 selects class probability values for only those
classes in which demographic D1 has participated. Recalling that
the example class Cartesian coordinates are in a binary
representation of (D1, D2), then the example linear combiner 122
only selects class probability values that include the demographic
of interest (e.g., D1). As such, the example linear combiner 122
selects values associated with Class 2 (29%) and Class 3 (71%) to
perform a summation (100%) for the probability that a household
member fitting demographic D1 participated in the corresponding
viewing behaviors.
[0022] In the event that one or more additional demographics of
interest for the household of interest are to be evaluated, the
example linear combiner 122 selects another demographic of interest
(e.g., demographic D2). In this case, only Class 1 (0%) and Class 3
(71%) meet the criteria for activity associated with demographic
D2. The example linear combiner 122 performs a summation of these
available values to calculate a 71% probability that a household
member fitting demographic D2 participated in the corresponding
viewing behaviors. When all demographics for the household of
interest have been evaluated, the example linear combiner 122
determines if one or more households of interest are to be
similarly evaluated, and the aforementioned calculations repeat as
described above.
[0023] As described above, the model is now trained and ready to be
applied to one or more test sets of data to determine predicted
probabilities when demographic information associated with viewing
behavior is unknown. In some examples, the test sets of data are
not associated with rigorous control and management typically
associated with panelist data. In some examples, the test sets of
data are associated with third party models, the consistency and/or
accuracy of which is unknown. Examples disclosed herein evaluate
such third party models in view of the training data to, in part,
determine whether the third party models are consistent with known
truth data and/or determine whether such third party truth models
should be ignored, accepted or accepted in part.
[0024] While the illustrated example of FIG. 2 included demographic
data 208 in which demographic values are a binary indicator (e.g.,
0 or 1) indicative of the absence or presence of a demographic of
interest, examples disclosed herein are not limited thereto. In
some examples, included demographic data may include values
indicative of a number of household members of a particular
demographic group that participated in the viewing behavior.
[0025] FIG. 3 illustrates example training data 300 retrieved by
the example data retriever 110 (e.g., from the example truth data
store 106) that, similar to the illustrated example of FIG. 2, is
segregated by five (5) example features of interest 302. Similar to
the illustrated example of FIG. 2, the illustrated example of FIG.
3 includes demographic data 308 with two separate demographic types
of interest; "D1" 310 and "D2" 312. However, unlike the binary
representation of demographics shown in the illustrated example of
FIG. 2, the values associated with the example demographic data 308
of FIG. 3 are indicative of a quantity of members within the
respective households that meet the demographic criteria of
interest. For example, in an example first household 306, D1 has a
value of "1" to indicate one household member fitting the
demographic criteria of D1, and D2 has a value of "2" to indicate
two household members fitting the demographic criteria of D2.
[0026] In the illustrated example of FIG. 3, uniquely occurring
combinations of demographic D1 and demographic D2 can include
between zero and three household members. As such, there are a
total of sixteen possible combinations of household members for D1
and D2 (two demographic values, each with four possible
values.fwdarw.2{circumflex over ( )}4=16). However, because
examples disclosed herein do not consider the combination where
zero household members were present for D1 and D2, a total of
fifteen (15) combinations exist with the illustrated example of
FIG. 3. An example class column 350 enumerates each uniquely
occurring combination of D1 and D2 as a separate class, beginning
with a designating letter "A" (Class A) to represent circumstances
in which household combinations have one member of type D1 and two
members of type D2 (e.g., (1,2)). While the unique combination of
Class A only occurs once in the illustrated example of FIG. 3 for
the example first household 306, an example Class E occurs five (5)
times in household 5, household 10, household 13, household 20 and
household 23. In particular, Class E reflects circumstances in
which there are three household members that satisfy the
demographic criteria of D1 and D2 (e.g., (3,3)). After considering
all unique combinations of D1 and D2 in the illustrated example of
FIG. 3, there are thirteen unique combinations (labeled as Class A
through Class M). However, considering that there is a potential
for fifteen unique combinations of D1 and D2 (in view of zero to
three members possible in each demographic of interest), the
example data 300 did not include the missing combinations of (2,1)
(e.g., two household members associated with D1 and one household
member associated with D2) and (3,0) (e.g., three household members
associated with D1, and no household members associated with
D2).
[0027] Despite the aforementioned differences regarding binary
demographic representations in the illustrated example of FIG. 2,
and numeric counts of household members for each demographic of
interest in the illustrated example of FIG. 3, the above identified
examples to develop a training model may be accomplished in
substantially the same manner. With the example model coefficients,
such as the example Class 1 model coefficient A (222), the example
Class 2 model coefficient B (224), and the example Class 3 model
coefficient C (226) of FIG. 2, one or more test sets may be applied
thereto.
[0028] In operation, the example data retriever 110 retrieves test
set data, such as observed behavior data for which corresponding
demographic information is unknown. In the illustrated example of
FIG. 2, the data retriever 110 retrieved a first test set 270
(e.g., from the example test data store 108) of quarter hours of
viewing time of the five (5) features of interest 202. The example
modeling engine 120 applies the model, including the example three
model coefficients A (222), B (224) and C (226) to derive
probability values for each class of interest, as shown in the
example test set results matrix 272. The example test set results
matrix 272 includes the three classes described above, which
includes a Class 1 cell 274, a Class 2 cell 276 and a Class 3 cell
278. While the illustrated example test set results matrix 272 also
includes a placeholder for D1=0 and D2=0, such circumstances do not
contribute to this analysis because it does not reveal behavior
information.
[0029] The result of the example modeling engine 120 performing the
multinomial regression on the example first test set 270 in
connection with the example three model coefficients A (222)
(associated with Class 1), B (224) (associated with Class 2) and C
(226) (associated with Class 3) are probability values of zero,
0.913 and 0.9087, respectively. As described above, these
probability values reflect class probabilities that include
component influences from the demographics of interest (D1 and D2).
The example class enumerator 116 reverts the classes back to
probability values for each demographic of interest by summing the
probability values of the test results matrix 272 according to each
demographic. In other words, because the classes include two or
more individual demographic components together (e.g., as a
probability value), the class enumerator 116 reverts those classes
into individualized demographic components, each having an
associated probability value. For example, cells of the example
test results matrix 272 that are associated with D1 behaviors
include the Class 2 cell (276) and the Class 3 cell (278). The sum
of resulting probability values of these cells adds up to 1.00,
which reveals that there is a 100% probability that the test set
data 270 behaviors included a household member belonging to
demographic D1. Similarly, cells of the test results matrix 272
that are associated with D2 behaviors include the Class 1 cell
(274) and the Class 3 cell (278). The sum of resulting probability
values of these cells adds up to 0.9087, which reveals that there
is a 91% probability that the test set data 270 behaviors included
a household member belonging to demographic D2.
[0030] As described above, examples disclosed herein also generate
models in which a training set includes third party models. For
example, Experian.RTM. collects information on people, businesses,
motor vehicles, insurance and lifestyle data (e.g., via online
surveys, offline surveys, social media monitoring, etc.). In some
examples, the third party models (e.g., from Experian.RTM.) exhibit
particular strengths to reveal demographic information associated
with particular demographics of interest. However, in some
examples, the third party models may not align well or otherwise
satisfy statistical requirements to allow predictions, projections
and/or estimations.
[0031] FIG. 4 illustrates example training data 400 retrieved by
the example data retriever 110 that, similar to the illustrated
examples of FIGS. 2 and 3, is segregated by five (5) example
features of interest 402. Similar to the illustrated examples of
FIGS. 2 and 3, the illustrated example of FIG. 4 includes
demographic data 408A with two separate demographic types of
interest associated with the training data 400; "D1" 410A and "D2"
412A. However, the illustrated example of FIG. 4 also includes a
third party model, which includes demographic data 408B with
representations of the same two separate demographic types of
interest; "D1" 410B and "D2" 412B. In some examples, third party
model data may be stored in the example test data store 108.
[0032] As described above, the example class enumerator 116
enumerates each unique demographic combination to a class
representation, and the example Cartesian combiner 118 assigns
Cartesian products of candidate combinations. In the illustrated
example of FIG. 4, nine (9) separate Cartesian products of
candidate combinations 414 are shown having coordinates in a format
(D1,D2) with a binary indicator of one (1) or zero (0) to indicate
the presence or absence, respectively, of household members
associated with demographic D1 and/or D2. Additionally, because the
illustrated example of FIG. 4 considers a comparison to a third
party model, Cartesian coordinate intersections of interest between
the truth model and the third party model are analyzed. For
example, a first intersecting column of interest 450 includes an
intersection of the third party model in which D1 does not
participate and D2 does participate in the viewing behavior (0,1)
with the truth model in which D1 does not participate and D2 does
participate with the viewing behavior (0,1). Taken together, the
first intersecting column of interest 450 is shown as "(0,1)
(0,1)." Additional columns of interest are shown in the illustrated
example of FIG. 4 for the remaining candidate combinations of
interest between the truth model and the third party model, and
there are nine (9) columns in total because circumstances where
there is no viewing behavior for D1 and D2 are not considered in
this analysis (e.g., (0,0) (0,1), (0,0) (1,0), (1,1) (0,0),
etc.).
[0033] As described above, the example modeling engine 120 performs
a multinomial logistic regression of the enumerated class
representations in connection with the training data 400 and the
third party model to generate probability splits for each class
representation. In the illustrated example of FIG. 4, the modeling
engine 120 calculates a 53% probability that the combination (0,1)
(0,1) occurs in connection with the associated training data 400.
In other words, there is a 53% probability that (a) the third party
model predicts that zero household members associated with D1 and
one household member associated with D2 participate in the viewing
behavior and (b) the truth model predicts the same.
[0034] In addition to calculating probability values for the class
intersections, the example modeling engine 120 also calculates
coefficients for each class of interest to be used for estimations
in the event one or more test data sets are obtained. However,
examples disclosed herein also reveal valuable information related
to the third party model regarding its ability to accurately
predict particular types of demographic behavior. As described
above, the example linear combiner 122 applies linear combinations
of the calculated class probabilities to estimate a likelihood that
a particular demographic is present. In the illustrated example of
FIG. 4, results of demographic likelihoods for the third party
model and the truth model 452 are shown. These example results 452
illustrate particular strengths or weaknesses of the third party
model for predicting demographic behaviors for particular
demographics of interest. For example, an example second household
413 illustrates that the modeling engine 120 and linear combiner
122 have calculated the third party model to predict a 69%
probability for D1 and a 61% probability for D2. Additionally, the
example modeling engine 120 and the linear combiner 122 have
calculated the truth model to predict a 71% probability for D1 and
a 63% probability for D2. The example linear combiner 122 compares
these predictions to determine difference values and, in this case,
the two models differ by 2% for D1 and 2% for D2. In some examples,
the linear combiner 122 compares these differences to a threshold
value that, when satisfied (e.g., exceeded), generates a flag to
warn against and/or otherwise prevent the third party model from
being relied upon for model correction or use. In other words, the
example linear combiner 122 facilitates an ability to generate a
trust metric for the third party model.
[0035] To illustrate, consider an example fourth household 454 that
illustrates the third party model predicts a 57% probability for
D1, while the truth model predicts a 25% probability for D1. In
this example, the third party model overpredicted demographic D1 by
32%. In the event a difference threshold were established at, for
example, 10%, then the example linear combiner 122 would flag the
third party model as unreliable for these circumstances.
[0036] However, in some examples additional information associated
with the third party model is available to reveal that a particular
demographic combination was empirically observed. In the event that
the third party model also includes observation data (e.g.,
empirical observation data) for a particular demographic
combination (e.g., D1=1, D2=0), then the example modeling engine
120 adjusts the truth model to align closer to the third party
model (e.g., weighting).
[0037] While an example manner of implementing the demographic
modeling system 100 of FIG. 1 is illustrated in FIG. 1, one or more
of the elements, processes and/or devices illustrated in FIG. 1 may
be combined, divided, re-arranged, omitted, eliminated and/or
implemented in any other way. Further, the example truth data store
104, the example test data store 106, the example data retriever
110, the example feature set segregator 112, the example household
data associator 114, the example class enumerator 116, the example
Cartesian combiner 118, the example modeling engine 120, the
example linear combiner 122, the example condition engine 124
and/or, more generally, the example demographic modeling manager
102 of FIG. 1 may be implemented by hardware, software, firmware
and/or any combination of hardware, software and/or firmware. Thus,
for example, any of the example truth data store 104, the example
test data store 106, the example data retriever 110, the example
feature set segregator 112, the example household data associator
114, the example class enumerator 116, the example Cartesian
combiner 118, the example modeling engine 120, the example linear
combiner 122, the example condition engine 124 and/or, more
generally, the example demographic modeling manager 102 of FIG. 1
could be implemented by one or more analog or digital circuit(s),
logic circuits, programmable processor(s), application specific
integrated circuit(s) (ASIC(s)), programmable logic device(s)
(PLD(s)) and/or field programmable logic device(s) (FPLD(s)). When
reading any of the apparatus or system claims of this patent to
cover a purely software and/or firmware implementation, at least
one of the example truth data store 104, the example test data
store 106, the example data retriever 110, the example feature set
segregator 112, the example household data associator 114, the
example class enumerator 116, the example Cartesian combiner 118,
the example modeling engine 120, the example linear combiner 122,
the example condition engine 124 and/or, more generally, the
example demographic modeling manager 102 of FIG. 1 is/are hereby
expressly defined to include a tangible computer readable storage
device or storage disk such as a memory, a digital versatile disk
(DVD), a compact disk (CD), a Blu-ray disk, etc. storing the
software and/or firmware. Further still, the example demographic
modeling system 100 of FIG. 1 may include one or more elements,
processes and/or devices in addition to, or instead of, those
illustrated in FIG. 1, and/or may include more than one of any or
all of the illustrated elements, processes and devices.
[0038] Flowcharts representative of example machine readable
instructions for implementing the demographic modeling system 100
of FIG. 1 are shown in FIGS. 5-8. In these examples, the machine
readable instructions comprise a program for execution by a
processor such as the processor 912 shown in the example processor
platform 900 discussed below in connection with FIG. 9. The program
may be embodied in software stored on a tangible computer readable
storage medium such as a CD-ROM, a floppy disk, a hard drive, a
digital versatile disk (DVD), a Blu-ray disk, or a memory
associated with the processor 912, but the entire program and/or
parts thereof could alternatively be executed by a device other
than the processor 912 and/or embodied in firmware or dedicated
hardware. Further, although the example program is described with
reference to the flowcharts illustrated in FIGS. 5-8, many other
methods of implementing the example demographic modeling system 100
may alternatively be used. For example, the order of execution of
the blocks may be changed, and/or some of the blocks described may
be changed, eliminated, or combined.
[0039] As mentioned above, the example processes of FIGS. 5-8 may
be implemented using coded instructions (e.g., computer and/or
machine readable instructions) stored on a tangible computer
readable storage medium such as a hard disk drive, a flash memory,
a read-only memory (ROM), a compact disk (CD), a digital versatile
disk (DVD), a cache, a random-access memory (RAM) and/or any other
storage device or storage disk in which information is stored for
any duration (e.g., for extended time periods, permanently, for
brief instances, for temporarily buffering, and/or for caching of
the information). As used herein, the term tangible computer
readable storage medium is expressly defined to include any type of
computer readable storage device and/or storage disk and to exclude
propagating signals and to exclude transmission media. As used
herein, "tangible computer readable storage medium" and "tangible
machine readable storage medium" are used interchangeably.
Additionally or alternatively, the example processes of FIGS. 5-8
may be implemented using coded instructions (e.g., computer and/or
machine readable instructions) stored on a non-transitory computer
and/or machine readable medium such as a hard disk drive, a flash
memory, a read-only memory, a compact disk, a digital versatile
disk, a cache, a random-access memory and/or any other storage
device or storage disk in which information is stored for any
duration (e.g., for extended time periods, permanently, for brief
instances, for temporarily buffering, and/or for caching of the
information). As used herein, the term non-transitory computer
readable medium is expressly defined to include any type of
computer readable storage device and/or storage disk and to exclude
propagating signals and to exclude transmission media. As used
herein, when the phrase "at least" is used as the transition term
in a preamble of a claim, it is open-ended in the same manner as
the term "comprising" is open ended.
[0040] The program 500 of FIG. 5 begins at block 502, where the
example data retriever 110 retrieves and/or otherwise receives
training data from the example truth data store 106. Depending on
the number and/or type of features of interest to be analyzed, the
example feature set segregator 112 segregates the retrieved
training data based on those feature sets of interest (block 504).
In some examples, one or more feature sets are selected to help
reveal audience behavior(s) and/or trends related to broadcast
assets, such as particular television shows or particular
commercials, etc. For example, the marketing analyst and/or a
client of the marketing analyst may be interested in learning about
which demographic household members tend to consume and/or
otherwise view the particular television programs broadcast on
media devices to a greater or lesser degree. For those particular
programs (e.g., features), the example feature set segregator 112
identifies particular households from the training set that include
viewing activity associated with the feature sets of interest, and
includes viewing data associated therewith (block 506). As such,
the viewing data to be included in the analysis is focused on (a)
the feature(s) of interest and (b) the viewing data for which those
features have occurred. In some examples, the viewing data includes
a number of quarter-hours of viewing time, as shown in FIGS.
2-4.
[0041] The segregated viewing data from the example truth data
store 106 is panelist data that also includes verified and/or
otherwise true demographic information. The example household data
associator 114 associates available demographic data with each
household of interest (block 508), which is shown as item 208 in
FIG. 2. The example class enumerator 116 enumerates each unique
demographic combination to a class representation (block 510). For
example, a combination of a household member associated with
demographic D1 not viewing and a household member associated with
demographic D2 viewing is enumerated as "Class 1." As another
example, a combination of a household member associated with
demographic D1 viewing and another household member associated with
demographic D2 viewing is enumerated as "Class 3." The enumerated
classes are assigned, by the example Cartesian combiner 118,
Cartesian product representations of candidate combinations (block
512). Examples of such enumerated assignments are shown as item 214
of FIG. 2.
[0042] The example modeling engine 120 performs a conditional
maximum entropy modeling of the truth data and enumerated
combinations by performing a logistic regression (e.g., a
multinomial logistic regression) of the enumerated class
representations (block 514). In particular, the example modeling
engine 120 generates probability splits for each class
representation for all participating households, and verifies that
all household probability values add up to 100%. Additionally, the
example modeling engine 120 fits the multinomial regression to
generate model coefficients associated with each unique enumerated
class (block 516), such as the example Class 1 model coefficient
"A" (222), the example Class 2 model coefficient "B" (224), and the
example Class 3 model coefficient "C" (226) described above in
connection with FIG. 2. As described above and in further detail
below, the example model coefficients may be used when generating
estimates in connection with test sets of data for which
demographic information is unknown.
[0043] While the example model class probabilities resulted from
the application of the multinomial regression, such probabilities
are not yet in terms of the demographic of interest, but rather two
or more demographic combinations that may have occurred with the
training (truth) data set (e.g., the example training data 200 of
FIG. 2, the example training data 300 of FIG. 3, the example
training data 400 of FIG. 4, etc.). To estimate the likelihood that
a particular demographic is present, the example linear combiner
122 applies linear combinations of the class probabilities that
represent a target demographic of interest (block 518).
[0044] FIG. 6 illustrates additional detail in connection with
applying a linear combination of interest of block 518. In the
illustrated example of FIG. 6, the example linear combiner 122
selects a household of interest (block 602) having associated class
probability values. To illustrate, examples from FIG. 2 are
described, but similar processes may be applied to the illustrated
examples of FIGS. 3 and/or 4. Assuming for this example that a
predicted probability for demographic D1 is desired (block 604),
the linear combiner 122 selects values from only those enumerated
classes in which D1 occurs (block 606). For example, Class 1
includes a Cartesian representation (0,1), Class 2 includes a
Cartesian representation (1,0), and Class 3 includes a Cartesian
representation (1,1). For each example Cartesian representation,
the first value reflects the occurrence or non-occurrence of
demographic D1, and the second value reflects the occurrence or
non-occurrence of demographic D2. For the sake of this example, in
which D1 is the target demographic of interest, the example linear
combiner 122 selects only Class 2 and Class 3 because they each
include potential behavior data associated with D1 (e.g., those
classes include a value of "1" in the Cartesian representation for
D1). In the illustrated example of FIG. 2, the sum of the values
for Class 2 (0.29) and Class 3 (0.71) is 1.00, which reflects a
100% probability that a household member associated with D1
participated in the activity (e.g., viewing behavior).
[0045] Values for those classes of interest are summed (block 608)
to yield a predicted probability for the demographic of interest.
The example linear combiner 122 determines whether an additional
demographic of interest is available for the selected household
(block 610). If so, control returns to block 604 to select an
alternate/additional demographic of interest. If not, then the
linear combiner 122 determines whether there are additional
households left to evaluate (block 612). If so, then control
returns to block 602, otherwise the example program 518 ends and
control returns to block 520 of FIG. 5.
[0046] As described above, after the training data has been modeled
to generate coefficients associated with each unique enumerated
class of demographic combinations, one or more test sets of data
may be applied to the model to estimate corresponding demographic
probabilities associated with those test set(s). In some examples,
a test set may include viewing behavior for features of interest,
such as quantities of viewed quarter hours of particular television
shows. In some examples, application of the test set occurs after
the training data has been modeled in connection with a third party
model, as shown in the illustrated example of FIG. 4. In some
examples, it is desired to determine whether the third party model
can be trusted, which may include determining particular
demographics of interest that the third party model estimates
relatively well as opposed to particular demographics of interest
that the third party model does not estimate well.
[0047] The example demographic modeling manager 102 determines
whether to evaluate a third party model (block 520) and, if so, the
example modeling engine evaluates the third party model (block 522)
on a demographic-by-demographic basis to ascertain how well or how
poorly it performs when compared to the truth model (which is based
on panelist-based training data). FIG. 7 illustrates additional
detail to evaluate the third party model in connection with block
522. In the illustrated example of FIG. 7, the example linear
combiner 122 compares probability results between the third party
model and the training model for demographic of interest (block
702). Depending on whether the third party model also has actual
(e.g., empirical) data associated with the demographic of interest,
examples disclosed herein will either use the third party model
results to help correct the training model or not. The example data
retriever 110 determines whether actual truth data is available
from the third party model for the demographic of interest (block
704) and, if so, the example modeling engine 120 adjusts the
training model with a weighted result of the probability result of
the third party model (block 706). In some examples, the weighted
result is based on the probability difference value between the
third party model and the training model.
[0048] On the other hand, if there is no trust data available from
the third party model associated with the demographic of interest
(block 704), then the example linear combiner determines whether
the deviation (difference) satisfies (e.g., exceeds) a threshold
value (block 708). In some examples, if the difference in
probability values estimated by the third party model and the
training model exceeds the threshold (e.g., 10%), then the third
party model validity and/or capability is deemed less reliable
(block 710). However, if the difference in probability values (for
the demographic of interest) estimated by the third party model and
the training model does not exceed the threshold (block 708), then
the example modeling engine 120 flags the third party model as
reliable (block 712). The linear combiner determines whether
additional demographics of interest are to be compared (block 714)
and, if so, control returns to block 702.
[0049] Returning to the illustrated example of FIG. 5, if there is
no third party model to be evaluated (block 520), or after such a
third party model has been evaluated (block 522), the example
demographic modeling manager 102 applies one or more test sets of
data to the training model (block 524). FIG. 8 illustrates
additional detail of applying test sets in connection with block
524. In the illustrated example of FIG. 8, the example data
retriever 110 retrieves test data (block 802), such as the example
first test set 270 of quarter hours of viewing time shown in FIG.
2. The example modeling engine 120 applies the training model
(e.g., the example Class 1 coefficient "A" 222, the example Class 2
coefficient "B" 224, and the example Class 3 coefficient "C" 226 of
FIG. 2) to the test set to derive probability values for each class
(block 804). As described above, classes include combinations of
demographic members that can later be reverted to individualized
demographic probabilities. The example class enumerator 116 reverts
the estimated class probabilities back to individualized
demographic probabilities for each demographic of interest (block
806). Control then returns to the illustrated example of FIG. 5
where the process 500 can repeat, as desired.
[0050] FIG. 9 is a block diagram of an example processor platform
900 capable of executing the instructions of FIGS. 5-8 to implement
the demographic modeling system 100 of FIG. 1. The processor
platform 900 can be, for example, a server, a personal computer, a
mobile device (e.g., a cell phone, a smart phone, a tablet such as
an iPad.TM.), a personal digital assistant (PDA), an Internet
appliance, a gaming console, a set top box, or any other type of
computing device.
[0051] The processor platform 900 of the illustrated example
includes a processor 912. The processor 912 of the illustrated
example is hardware. For example, the processor 912 can be
implemented by one or more integrated circuits, logic circuits,
microprocessors or controllers from any desired family or
manufacturer. In the illustrated example of FIG. 9, the processor
900 includes one or more example processing cores 915 configured
via example instructions 932, which include the example
instructions of FIGS. 5-8 to implement the example demographic
modeling system 100 of FIG. 1.
[0052] The processor 912 of the illustrated example includes a
local memory 913 (e.g., a cache). The processor 912 of the
illustrated example is in communication with a main memory
including a volatile memory 914 and a non-volatile memory 916 via a
bus 918. The volatile memory 914 may be implemented by Synchronous
Dynamic Random Access Memory (SDRAM), Dynamic Random Access Memory
(DRAM), RAMBUS Dynamic Random Access Memory (RDRAM) and/or any
other type of random access memory device. The non-volatile memory
916 may be implemented by flash memory and/or any other desired
type of memory device. Access to the main memory 914, 916 is
controlled by a memory controller.
[0053] The processor platform 900 of the illustrated example also
includes an interface circuit 920. The interface circuit 920 may be
implemented by any type of interface standard, such as an Ethernet
interface, a universal serial bus (USB), and/or a PCI express
interface.
[0054] In the illustrated example, one or more input devices 922
are connected to the interface circuit 920. The input device(s) 922
permit(s) a user to enter data and commands into the processor
1012. The input device(s) can be implemented by, for example, an
audio sensor, a microphone, a camera (still or video), a keyboard,
a button, a mouse, a touchscreen, a track-pad, a trackball,
isopoint and/or a voice recognition system.
[0055] One or more output devices 924 are also connected to the
interface circuit 920 of the illustrated example. The output
devices 924 can be implemented, for example, by display devices
(e.g., a light emitting diode (LED), an organic light emitting
diode (OLED), a liquid crystal display, a cathode ray tube display
(CRT), a touchscreen, a tactile output device, a printer and/or
speakers). The interface circuit 920 of the illustrated example,
thus, typically includes a graphics driver card, a graphics driver
chip or a graphics driver processor.
[0056] The interface circuit 920 of the illustrated example also
includes a communication device such as a transmitter, a receiver,
a transceiver, a modem and/or network interface card to facilitate
exchange of data with external machines (e.g., computing devices of
any kind) via a network 926 (e.g., an Ethernet connection, a
digital subscriber line (DSL), a telephone line, coaxial cable, a
cellular telephone system, etc.).
[0057] The processor platform 900 of the illustrated example also
includes one or more mass storage devices 928 for storing software
and/or data. Examples of such mass storage devices 928 include
floppy disk drives, hard drive disks, compact disk drives, Blu-ray
disk drives, RAID systems, and digital versatile disk (DVD) drives.
In some examples, the mass storage device may implement the example
truth data store 106 and/or the example test data store 108.
[0058] The coded instructions 932 of FIGS. 5-8 may be stored in the
mass storage device 928, in the volatile memory 914, in the
non-volatile memory 916, and/or on a removable tangible computer
readable storage medium such as a CD or DVD.
[0059] From the foregoing, it will be appreciated that the above
disclosed methods, apparatus and articles of manufacture improve
the efficiency when modeling multiple demographics of interest from
training data so that estimates may be generated from one or more
test sets of data. In particular, traditional approaches to model a
first demographic of interest include a first modeling iteration in
view of only that particular demographic of interest. Once that
modeling effort has been completed, a subsequent modeling effort
(e.g., using the same training data) is performed on the alternate
demographic of interest. This process repeats for all demographics
of interest that are of interest and, when corresponding
coefficients for the demographics of interest have been calculated
(e.g., fitted), further modeling efforts must be applied to
compensate for potential modeling errors caused by co-linearity.
However, examples disclosed herein reduce (e.g., minimize) a number
of modeling iterations for training data in which two or more
demographics of interest are to be estimated. Additionally,
examples disclosed herein reduce modeling errors in connection with
reliance upon third party models, in which particular estimation
and/or prediction capabilities are unknown. Despite an analyst
belief or "word of mouth" that a particular third party model is
helpful for demographic research, overreliance on such models when
they fail to produce accurate estimates causes wasted research
efforts and computational waste.
[0060] Although certain example methods, apparatus and articles of
manufacture have been disclosed herein, the scope of coverage of
this patent is not limited thereto. On the contrary, this patent
covers all methods, apparatus and articles of manufacture fairly
falling within the scope of the claims of this patent.
* * * * *