Identifying Geographical Areas using Machine Learning for Enrolling

Women in the Canadian Armed Forces

Ryuichi Ueno, Peter Boyd

and Dragos Calitoiu

Department of National Defence, 60 Moodie Drive, Ottawa, ON, K1A 0K2, Canada

Keywords:

Feature Selection, Clustering, Logistic Regression, Propensity Scores.

Abstract:

To improve the visibility of military service as a career option for women, the Canadian Armed Forces (CAF)

can tailor marketing campaigns to geographical areas and demographics within Canada that have historically

high enrollment of women. To aid in this recruitment strategy, a logistic regression model was developed

using historical recruiting data. The score obtained was used to rank Canadian postal codes and to identify the

ones with the highest potential for recruiting of women. Additional demographic ﬁltering was applied using

marketing segments provided by a vendor. The ﬁnal top 10% postal codes with the highest probability of

women enrollment were clustered based on the collective social media behaviour of each postal code and was

binned using the distance to the nearest recruiting centre. Several social media outlets were observed to be of

interest, among them YouTube and Snapchat appear as viable options to reach women with a high probability

of CAF enrollment.

1 INTRODUCTION

The Canadian Armed Forces (CAF) is committed to

increasing the number of women from 15.9%

25.1%, as strategically emphasized in the latest de-

fence policy (Strong, Secure, Engaged) (Department

of National Defence, 2017). In order to rapidly in-

crease the enrolment of women, tailored recruitment

strategies must be used, focusing on attraction and se-

lection of suitable women candidates for military ser-

vice within speciﬁc military occupations. Efforts that

are customized to different subgroups within a pop-

ulation based on past success can also contribute to

increasing enrollment.

Ueno et al. applied unsupervised machine learn-

ing to group Canadian postal codes into clusters, us-

ing demographic information as features to partition

the population (Ueno et al., 2019). The clusters were

then ranked based on historical application and enroll-

ment data. A ranking scheme was devised that con-

sidered both successful and unsuccessful applicants

(men and women), making it adaptable to speciﬁc

campaign requirements.

In the United States (U.S.), a similar approach of

identifying geographical locations with potential for

https://orcid.org/0000-0001-5023-1200

https://orcid.org/0000-0003-0173-9846

As of the end of 2019.

recruiting has been recently explored (Buttrey et al.,

2018; Fulton, 2016; Monaghan, 2016). The U.S.

Army uses commercial market segmentation data to

analyze markets and past enrollment to assign re-

cruiters and quotas to maximize production. Ful-

ton (Fulton, 2016) proposed a clustering approach

using ZIP codes, exploiting the different densities

of potential recruits. Fulton’s approach considered

347 variables from publicly available U.S. govern-

ment agencies for all 34,007 ZIP codes to cluster

them into similar groups. By using three dissimilarity

measures and three clustering algorithms, he identi-

ﬁed 18 clusters, “a new method of identifying poten-

tially high production ZIP codes” and “an easier tool

for recruiting commanders” (Fulton, 2016). Mon-

aghan (Monaghan, 2016) expanded Fulton’s work to

the U.S. Navy by developing ﬁve-cluster membership

factors based on publicly available data sources, for

helping Navy Recruiting Command predict the num-

ber of leads as an indicator of the market depth that a

ZIP code will produce.

The U.S. Army Recruiting Command (USAREC)

uses a segmentation developed by Nielsen (former

Claritas): Potential Rating Index for Zip Markets

(PRIZM) (Nielsen Company, 2020). Each house-

hold is assigned to one of the 68 segments, describing

demographic and consumer behaviour. Since 2007,

USAREC augmented the PRIZM segments with atti-

Ueno, R., Boyd, P. and Calitoiu, D.

Identifying Geographical Areas using Machine Learning for Enrolling Women in the Canadian Armed Forces.

DOI: 10.5220/0010186703070316

In Proceedings of the 10th International Conference on Operations Research and Enterprise Systems (ICORES 2021), pages 307-316

ISBN: 978-989-758-485-5

307

Figure 1: Overall ﬂow of the method presented in this paper. Each of the steps are detailed in the respective sections in the

text.

tudes, motivators and barrier to service data and de-

veloped 39 Army Custom Segments (Clingan, 2007),

called tactical segments. USAREC also grouped them

into 10 strategic segments and determined which seg-

ments have a population with a higher propensity for

joining the military. Recruiters used this information

to better decide who they should attempt to recruit and

how.

PRIZM has not only been used for the purposes

of military recruiting, these demographic categories

are used predominately in the business industry for

targeted marketing. Aside from these sectors, they

are also used in a wide variety of ﬁelds to identify

geographical and demographical dimensions of a par-

ticular attribute. For example, in retailing and man-

ufacturing, PRIZM segments were used to predict in

which geographical regions product afﬁnity would be

highest (Ayzenshtat et al., 2019). In urban planning,

demographics of past and present inhabitants were

found to dictate trends in urban landscape and veg-

etation characteristics (Boone et al., 2010). PRIZM

segments were used also to identify sensitive demo-

graphics that live near hazardous waste sites in the

US (Heitgerd, 2001).

The goal of this current research is to help re-

cruiters identify geographical areas where women are

more likely to join the CAF Regular Forces (RegF). In

order to do this, we ﬁrst used historical data, namely

the home address of the women who applied to be

hired as RegF members between January 2016 and

March 2019. For each applicant, their postal code and

a binary recruiting success variable (enrolled/not en-

rolled) were identiﬁed. A logistic regression model

was built to discriminate postal codes for recruiting

success. As independent variables, we explored 760

attributes that combine econometric, demographic

and geographic characteristics at the postal code level.

The logistic regression model produced a score that

was applied to all 853,109 Canadian postal codes. Us-

ing the score as a ranking measure, the top 10% of

postal codes were selected as the geographical loca-

tions of the women most likely to join the CAF.

The Canadian version of Nielsen’s PRIZM seg-

ments were used to identify the locations produc-

ing women enrollments to complement the statisti-

cal model developed in-house. This PRIZM dataset

contains 68 segments that capture demographics,

lifestyle, consumer behaviour, and settlement patterns

in Canada. The ﬁve most productive segments, cov-

ering 14% of the Canadian population, were able to

collect almost 25% of all women enrollments.

The selected postal codes were further segmented

to tailor to future recruiting campaigns. First, we used

collective social media behaviour of each postal code

to segment them and identify the best marketing ap-

proaches. Postal codes with similar social media be-

haviour were clustered using unsupervised machine

learning. Second, the distances from each postal code

to the nearest recruiting centre were computed, with

the knowledge that the cost and efﬁciency of the re-

cruiting process is correlated with the distance from

the selected postal code to the recruiting centre.

The novelty of our proposal resides in the design

of the process of searching for the best geographical

location. The process contains two steps: (i) the se-

lection, and (ii) the description of the selected loca-

tion using two measures. The selection is done by

combining an individual postal code score (obtained

with logistic regression) with a collective score (from

PRIZM segments

). The selection has these two sep-

arate scores because the PRIZM numbers are repre-

sented as categorical data and are not eligible to be in-

dependent variables in the logistic regression model.

With these selections, we can help marketers with the

description of the selection to identify the best com-

munication channel and message to the potential ap-

plicants. Two measures are used for describing the

selection: the social media behaviour and the distance

to the nearest recruiting centre. We do not eliminate

Compared to the collective score applied in the U.S.,

the Canadian PRIZM segments are derived at the postal

code level. This granularity allows for a better modeling;

the U.S. ZIP code is at least 10 times larger by population

count compared to the Canadian postal code.

ICORES 2021 - 10th International Conference on Operations Research and Enterprise Systems

308

more based on our description of the selection, but

instead we offer the marketers options to adapt their

action plan to the proposed selection.

Section 2 details the datasets used for this work.

Section 3 describes the logistic regression model.

Section 4 lays out the outcomes obtained by apply-

ing the PRIZM segments. Section 5 introduces the

clusters for customizing the marketing message. Sec-

tion 6 presents the binning considering the distance

from the home postal code of the applicant to the clos-

est recruiting centre, followed by conclusions in sec-

tion 7. See Figure 1 for the visual depiction of the

overall methodology.

2 DATASET

Four datasets were used for this paper, all of which

were obtained at the postal code granularity. Each of

the datasets is described below.

2.1 Women Applicant Dataset

This dataset was retrieved from the Canadian Forces

Recruiting Group (CFRG) and it contains women ap-

plicants to the CAF from January 2016 to March

2019. We removed applicants with no ﬁnal decision,

keeping only the enrollments and non-enrollments.

We retrieved postal codes from each applicant and

they were counted and labelled into 1,513 postal

codes with enrollments and 12,507 postal codes with

non-enrollments. If a postal code had both enroll-

ments and non-enrollments, it was labelled as as an

enrollment.

2.2 DemoStats

A demographic dataset called DemoStats is pro-

vided by Environics Analytics, a data, analytics

and marketing services company specializing in

geo-demographic segmentation (Environics Analyt-

ics, 2018). A total of 760 demographic attributes were

considered as potential candidates to explain the tar-

get variable, the enrollment. DemoStats is built from

a variety of data sources, including the latest cen-

sus, economic indicators, postcensal estimates from

the federal and provincial governments, immigration

statistics and economic data. It features variables

on population, family structure, household size and

type, ethnic diversity, labour force participation and

income.

2.3 PRIZM Segments

The PRIZM segments used for this paper are the

Canadian version of Nielsen’s PRIZM segments as

described in the previous section. This dataset was

also acquired from Environics Analytics, and is used

to identify the locations producing women recruits.

The PRIZM dataset contains 68 segments capturing

demographics, lifestyle, consumer behaviour and set-

tlement patterns across Canada. Since this dataset is

entirely categorical, it is used as a ﬁltering tool, not as

a dataset for modelling.

2.4 Opticks Social

The social media dataset consists of 720 social me-

dia attributes in Opticks Social provided by En-

vironics Analytics. These attributes are derived

from an online survey of 19,938 respondents from

AskingCanadians

, and describes social media us-

age, frequency and behaviours. The survey data are

matched to postal code by using the k-nearest neigh-

bour approach in conjunction with 50 Demostats at-

tributes and match the postal code with the smallest

differences.

3 LOGISTIC REGRESSION

MODEL (DEVELOPING AND

SCORING)

We have compared logistic regression model with

two neural network models (a multi-layer perceptron,

MLP, and a deep neural network, DNN) in a related

study on a dataset involving all recruits, and found

that the difference in performance was marginal for

previous dataset. The much smaller size of the train-

ing set used for this study (since it focuses on women)

makes it unsuitable for training neural networks and,

therefore, we opt to use logistic regression for scoring

the postal codes.

Logistic regression models the probability of be-

longing to a class A when there are only two mutu-

ally exclusive classes, namely class A (enrollment)

and class not-A (non-enrollment). Instead of using

a linear regression model to represent these probabil-

ities:

p(X

, X

, ...,X

) = β

+β

+... +β

(1)

where X

, X

, ..., X

are the dependent variables, we

modelled a function that gives outputs between 0 and

1 for all values of X

. Logistic regression does this by

using the logistic function:

Identifying Geographical Areas using Machine Learning for Enrolling Women in the Canadian Armed Forces

309

Table 1: The most signiﬁcant features found from logistic regression.

Name of the

Variable

Contribution

to the model

Pr>ChiSq Parameter

Estimation

Mean Min Max Std Cap VIF

intercept -2.36720

ECYINDPUBLP 74.81% <0.0001 0.00594 10.93 0 1366 39.43 129.22 1.00

ECYACTUR 9.84% <0.0001 -0.00206 7.19 0 62.5 6.48 26.63 1.01

ECYINDWHOL 6.86% 0.0043 0.00677 5.32 0 600 17.52 57.89 1.18

ECYACTPR 5.77% 0.0008 0.00615 66.84 0 100 15.47 N/A 1.00

ECYOCCSCND 2.73% 0.0016 0.00394 8.03 0 1209 29.95 97.88 1.20

p(X

, X

, ..., X

) =

exp[β

+ β

+ ... + β

]

1 + exp [β

+ β

+ ... + β

]

(2)

The logistic function always produces an S-shaped

curve, and regardless of the value of X

, we obtain

a sensible prediction.

For deriving the logistic regression model, Base

SAS version 9.4 was used. For the model devel-

opment, the women applicant dataset was combined

with the Demostats dataset using the postal code as

joining key. The dataset was randomly split 50%-50%

into development and validation sets. All Demostats

variables are numerical, positive real numbers. A cap-

ping on the right with 3σ was implemented, a stan-

dard procedure for assuring the stability of the model.

The cap values for the ﬁnal selection are presented in

Table 1. A combination of stepwise regression and

multicollinearity was used to reduce the number of

variables. SAS PROC LOGISTIC (EVENT=’0’) was

implemented for the stepwise selection, with a sig-

niﬁcance level of 0.1 (SLE=0.1) required to allow a

variable into the model, and a signiﬁcance level of

0.01 (SLS=0.01) required for a variable to stay in the

model. The Hosmer and Lemeshow goodness-of-ﬁt

test (Harnett, 1975) for the ﬁnal selected model was

requested by specifying the LACKFIT option.

74.81%

9.84%

6.86%

5.77%

2.73%

0% 10% 20% 30% 40% 50% 60% 70% 80%

ECYINDPUBLP

ECYACTUR

ECYINDWHOL

ECYACTPR

ECYOCCSCND

Figure 2: Variable contributions.

The contributions of the selected variables are pre-

sented in Figure 2. We report here only the parameters

associated to the variables that have a signiﬁcant con-

tribution to the logistic regression model: ECYIND-

PUBLP (Household Population 15 years or over by

Industry: Public Administration), ECYINDWHOL

(Household Population 15 years or over by Industry:

Wholesale Trade), ECYOCCSCND (Household Pop-

ulation 15 years or over by Industry: Occupations

unique to manufacture and utilities), ECYACTUR

(Household Population 15 years or over by Labour

Force Activity: Participation Rate), and ECYACTPR

(Household Population 15 years or over by Labour

Force Activity: Unemployment rate). The partici-

pation rate represents the percentage of people who

are in the labour force (namely employed or actively

seeking work); unemployment rate represents the per-

centage within the labour force that is currently with-

out a job.

The parameter estimations, namely the coefﬁ-

cients of the logistic regression for the target variable

being zero, are presented in Table 1. Two descrip-

tors regarding the possible multicollinearity (VIF

and Condition index

) are computed. SAS PROC

REG /vif and PROC REG /collin were applied for re-

moving collinearity. VIF is presented in Table 1. The

Condition index was smaller than 10. Each record re-

ceives a score, namely enrollment score, and we ap-

ply Equation 2 to map the score into a probability as

written below:

p 0 1 = 1 −

exp[enrollment score]

1 + exp [enrollment score]

(3)

The variance inﬂation factor (VIF) quantiﬁes the sever-

ity of multicollinearity in an ordinary least square regres-

sion analysis. It provides an index that measures how much

the variance (the square of the estimate’s standard devia-

tion) of an estimated regression coefﬁcient is increased be-

cause of collinearity. A rule of thumb is that if VIF > 3,

then multicollinearity is high.

The condition index is the standard measure of ill-

conditioning in a matrix. It is computed by ﬁnding the

square root of the maximum eigenvalue divided by the min-

imum eigenvalue. If the condition index is above 30, the

regression may have a signiﬁcant multicollinearity. Mul-

ticollinearity exists if two or more of the values related to

the high condition index have a high proportion of variance

explained. One advantage of this method is that it shows

which variables are collinear.

ICORES 2021 - 10th International Conference on Operations Research and Enterprise Systems

310

Table 2: Logistic regression results.

Decile

Development data Validation data

p 0 1 Enrollment

Enrollment

per decile

p 0 1 Enrollment

Enrollment

per decile

Mean Sum ColPctSum Mean Sum ColPctSum

1 0.1833 159 20.73 22.65% 0.1842 147 19.73 21.00%

2 0.1319 89 11.58 12.68% 0.1326 106 14.22 15.14%

3 0.1238 81 10.54 11.54% 0.1241 82 11.00 11.71%

4 0.1176 81 10.54 11.54% 0.1177 74 9.93 10.57%

5 0.1117 75 9.76 10.68% 0.1116 71 9.53 10.14%

6 0.1055 74 9.63 10.54% 0.1056 63 8.45 9.00%

7 0.0990 59 7.68 8.40% 0.0986 62 8.32 8.86%

8 0.0904 54 7.03 7.69% 0.0897 51 6.84 7.29%

9 0.0792 53 6.90 7.55% 0.0785 48 6.44 6.86%

10 0.0519 43 5.59 6.13% 0.0510 41 5.50 5.86%

We scored both the development and the validation

set, which consist of 702 and 700 postal codes, re-

spectively. A probability was computed from the

score, and then the postal codes were sorted in de-

scending order by the probability. We built deciles,

ranking each postal code by the probability to enroll

women and ordering from highest to lowest. We re-

ported for each decile the following numbers: the size

of the decile (number of postal codes), the mean of the

enrollment probability predicted by the model (p), the

number of enrollments realized on that decile (sum),

the percentage of the enrollments collected in that

decile out of total enrollments (ColPctSum) and, ﬁ-

nally, the enrollment rate per decile. Table 2 contains

these values for the development data and for the vali-

dation data. A model is powerful if it can group a high

percentage of all enrollments in the ﬁrst decile. For a

random selection, we expect to collect 10% of all en-

rollments in the ﬁrst decile. In our model, 20% of all

enrollments were collected in the ﬁrst decile, obtain-

ing a lift of 200% (model vs. random). If the model

has enrollment per decile descending from decile 1

to 10 without reversals (the stair case effect) then the

model is statistically correct. Table 2 supports the

lack of reversals in both the development and valida-

tion data.

In order to quantify the efﬁciency of the model,

the Kolmogorov-Smirnoff (KS) statistic is com-

puted

. For the development data set, KS=0.1559; for

the validation data set, KS=0.1346, both of them be-

The Kolmogorov-Smirnoff statistic is used to measure

the discriminatory power of a model, namely how the dis-

tribution of the score differs among enrollments and non-

enrollments. The KS measures the minimum point of sepa-

ration between the cumulative distribution functions of two

distributions.

ing greater than 0.1, an acceptable value.

We extend this model to score the entire 853,109

Canadian postal codes. A probability is computed

from the score, and then the postal codes are sorted

in descending order by the probability. We select the

top 10% for the next steps (85,310 postal codes).

4 FILTERING WITH THE PRIZM

SEGMENTS

Using all of the women applicant data and all of the

postal codes in Canada, it is found that the ﬁve most

productive PRIZM segments are able to collect al-

most 25% from all enrollments: PRIZM 32 (Mini Van

& Vin Rouge) 6.60%, PRIZM 37 (Trucks & Trades)

4.84%, PRIZM 16 (Pets & PCs) 4.23%, PRIZM 24

(Fresh Air Families) 4.23%, and PRIZM 63 (Lunch

at Tim’s) 3.75%.

Figure 3: Geographical locations of the postal codes in

PRIZM segment 32.

Out of 85,310 postal codes selected from the lo-

gistic regression, we identiﬁed 19,884 postal codes

from the ﬁrst ﬁve PRIZM segments. We ranked all the

Identifying Geographical Areas using Machine Learning for Enrolling Women in the Canadian Armed Forces

311

Table 3: Distribution of selected postal codes into PRIZM

segments by age group.

PRIZM Rank postal

codes

age

17–24

age

25–29

32 1 8,594 55,083 43,899

37 10 2,473 12,119 10,360

16 16 2,060 8,874 6,754

24 9 3,206 46,214 25,620

63 6 3,551 4,237 3,266

PRIZM segments in the ﬁrst decile to validate where

the ﬁve best PRIZM segments were distributed. The

ranks are presented in Table 3. This additional ﬁlter-

ing complements the statistical model and adds ex-

tra stability in the ﬁnal selection. The distribution of

these postal codes by PRIZM segment is presented

in Table 3, together with the women population con-

tained in these postal codes for three age groups: 17-

19, 20-24, and 25-29. These statistics are useful since

most of the recruits into the CAF are under age of

30 (Department of National Defence, 2018). The ge-

ographical distribution for PRIZM 32 is presented in

Figure 3 (The distribution for all ﬁve PRIZM seg-

ments and their descriptinos are presented in the Ap-

pendix.

5 CLUSTERING ON SOCIAL

MEDIA BEHAVIOUR

In order to assist marketing efforts for future recruits,

we segmented the 19,884 postal codes into several

clusters based on social media behaviours and made

recommendations on how to best market to these pop-

ulations. We employed an unsupervised machine

learning technique to blind the actual recruit informa-

tion, similar to the approach taken in (Ueno et al.,

2019).

For this purpose, we used the Opticks Social

dataset from Environics Analytics. This dataset was

ﬁrst standardized to remove means and to scale to

unit variance, then aggregated down to 72 features

using a Feature Agglomeration (FA) technique with

Ward linkage criterion that minimizes the variance of

the aggregated features (Ward, 1963). This method

is implemented in the Scikit-learn library version

0.22.1 with Python version 3.7.1. (Pedregosa et al.,

2011)

5.1 Developing Clusters

We used a Gaussian mixture model (GMM) to cluster

the social media dataset. GMM assumes the data con-

tains a ﬁnite number of clusters that are normally dis-

tributed, and uses an expectation-maximization (EM)

technique to ﬁt the mean and variance of the multi-

variate Gaussian distribution (Dempster et al., 1977).

GMM assigns nodes to several clusters probabilisti-

cally in the absence of sharp separation boundaries.

This technique is implemented in the Scikit-learn

package.

One of the main and difﬁcult problems in cluster-

ing is ﬁnding the right number of clusters to describe

the observations. For solving this problem, we used

the Bayesian Information Criterion (BIC) as deﬁned

by:

BIC = ln (n)k − 2ln





(4)

where

L is the maximized likelihood value, n is the

number of data points, and k is the number of clus-

ters. BIC will correct for over-ﬁtting of a model by

introducing a penalty for additional free parameters

for higher k.

Figure 4 shows the BIC obtained as a function

of the number of clusters, where the score for each

cluster conﬁguration is an average of 10 independent

trials. As can be seen, the 7-cluster conﬁguration

is favourable. However, two of the small clusters

were close in Euclidean distance; therefore they were

merged together, and six ﬁnal clusters were identiﬁed.

Figure 4: BIC scores of GMM clustering as a function of

number of clusters. The BIC score is an average of 10 inde-

pendent trials for each cluster conﬁguration.

Table 4: Result of GMM clustering.

Cluster

PRIZM 1 2 3 4 5 6

32 7,776 37 0 23 746 12

37 2,020 9 0 0 444 0

16 1,235 25 0 0 800 0

24 2,398 70 28 0 710 0

63 3,474 4 0 0 73 0

Total 16,903 145 28 23 2,773 12

Table 4 shows the result of the GMM clustering,

separated into the 5 PRIZM segments as presented be-

fore. As it can be seen, most of the postal codes be-

ICORES 2021 - 10th International Conference on Operations Research and Enterprise Systems

312

Table 5: Summary of social media (SM) behaviour by GMM cluster.

Feature

Cluster Importance Summary of behaviours

1 & 5 12.1 % - Current podcast listener, but not sports or technology focused podcast

- Uses Wikis to read articles or for research a few times a week

2 12.3 % - Typically uses mobile phone to access SM

- Watches YouTube frequently, but not active on Instagram

- Share content on SM forums and chats

3 11.3 % - Snapchat user, but not active on LinkedIn

- active on chats. Shares links with friends and colleagues

4 10.0 % - Typically uses computer to access SM

- Low use of YouTube, Instagram, or Facebook (a few times a month)

- Uses SM for work/professional purposes than personal use

6 15.1 % - Flickr user

long in cluster 1 and 5. Cluster 3 describes a small

subset of PRIZM segment 24 while Clusters 4 and 6

describe a small subset of PRIZM segment 32. Fig-

ure 5 shows the distribution of the probability from

the logistic regression model within each cluster. As

can be seen, the clusters generally correlate with the

probabilities from the logistic regression model and

are able to distinguish the postal codes with strong

probability from those with weaker probability.

Figure 5: Distribution of postal codes by the logistic re-

gression model probability. The colours correspond to the

cluster of each postal code.

5.2 Social Media Behaviour of Clusters

In order to guide marketing strategies, we proﬁled

each social media cluster identiﬁed above. In order

to do so, we employed a random forest classiﬁer (im-

plemented in the scikit-learn library) to determine

which features are most important for each of the six

clusters. Since the social media attributes are aggre-

gated into 72 groups of similar features, we were able

to proﬁle each cluster using several descriptions.

Table 5 shows the summary of the social media

behaviour of clusters. As can be seen, the clusters por-

tray distinct social media behaviour except for clus-

Table 6: Histogram of distances from postal codes in the

top 5 PRIZM categories to the nearest recruiting centre.

Distance (km)

PRIZM <25 25–50 50–75 75–100 >100

32 4,119 2,204 850 341 1,080

37 723 160 109 85 1,396

16 1,076 281 55 48 600

24 869 808 401 340 788

63 1,004 436 367 273 1,471

Total 7,791 3,889 1,782 1,087 5,335

ters 1 and 5, which are very close in terms of logistic

regression probability (See Figure 5). For example, in

order to reach the population in cluster 3, Snapchat is

preferable to LinkedIn. For cluster 2, YouTube adver-

tisement is a viable option for recruiting women.

6 BINNING BASED ON THE

DISTANCE TO THE NEAREST

RECRUITING CENTER

Interactions between recruiters and applicants vary

based on the distance between an applicant’s resi-

dence and the recruiting centre. Therefore, we com-

puted the distance from each of the selected 19,884

postal codes to the closest recruiting centre. There

are 26 recruiting centres in Canada. The distance

was computed between the center of the postal codes

of the recruiting centre and selected postal code de-

scribed by longitude and latitude coordinates. We

grouped the distance into ﬁve bins: less than 25 km,

between 25 and 50 km, between 50 and 75 km, be-

Identifying Geographical Areas using Machine Learning for Enrolling Women in the Canadian Armed Forces

313

tween 75 and 100 km, and more than 100 km. Al-

most 60% from all selected postal codes were less

than 50 km from the nearest recruiting centre. The

number of postal codes in each bin per PRIZM seg-

ment is presented in Table 6. This allows an addi-

tional tool for the marketers to focus on the recruiting

strategy based on the distance.

7 CONCLUSIONS

In the absence of the individual data for the potential

applicants, we explored the problem of identifying

geographical areas with high potential for recruiting

women by using demographic, life style, consumer

behaviour, settlement patterns and social media char-

acteristics of the neighbourhood (postal code) of the

potential applicants. Using historical women recruit-

ing data and two Environics datasets (DemoStats and

PRIZM), a selection of 19,884 postal codes with high-

est recruiting potential was obtained from the most

promising 10% of all Canadian postal codes, obtained

using a logistic regression model. Finally, the selec-

tion was segmented to derive the optimum marketing

channels and messages by using two approaches: (i)

clustering based on the aggregated social media be-

haviour of the postal code and (ii) binning using the

distance to the nearest recruiting centre. The social

media behaviour was retrieved from a third Environ-

ics dataset, Optiks Social. The approach presented in

this paper can be applied frequently for a long term,

contributing to the commitment to increase women

representation in CAF.

REFERENCES

Ayzenshtat, L., Rajamani, K., Vishnevskiy, A., Georgiev,

N., and Preotescu, M. (2019). Methods and appara-

tus to identify afﬁnity between segment attributes and

product characteristics. US PATENT 0073685 A1.

Boone, C., Cadenasso, M., Grove, J., Schwarz, K., and

Buckley, G. (2010). Landscape, vegetation charac-

teristics, and group identity in an urban and suburban

watershed: Why the 60s matter. Urban Ecosystems,

13(3):255–271.

Buttrey, S. E., Whitaker, L. R., and Alt, J. K. (2018). De-

velopments in the statistical modeling of military re-

cruiting. CHANCE, 31(2):38–44.

Clingan, M. (2007). U.s. army custom segmentation sys-

tem. Presentation at 75th Military Operations Re-

search Society Symposium, June 12-14, Annapolis,

MD.

Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977).

Maximum likelihood from incomplete data via the em

algorithm. Journal of the Royal Statistical Society, Se-

ries B, 39(1):1–38.

Department of National Defence (2017). Strong, secure,

engaged: Canada’s defence policy.

Department of National Defence (2018). Annual Report on

Regular Force Personnel:2015-2016. Internal Refer-

ence Document: DRDC-RDDC-2018-D086.

Environics Analytics (2018). Demostats database.

https://www.environicsanalytics.com/en-ca/data/

demographic/demostats. Accessed Feb. 20, 2019.

Fulton, B. M. (2016). Determining Market Categorization

of United States Zip Codes for Purposes of Army Re-

cruiting. Master’s thesis, Naval Postgraduate School,

Monterey, CA.

Harnett, D. L. (1975). Introduction to statistical methods.

Addison-Wesley Pub. Co.

Heitgerd, J. L. (2001). Using gis and demographics to char-

acterize communities at risk: A model for atsdr. Jour-

nal of Environmental health, 64(5):21–30.

Monaghan, E. M. (2016). Estimating the depth of the navy

recruiting market. Master’s thesis, Naval Postgraduate

School, Monterey, CA.

Nielsen Company (2020). My best segments. https://

claritas360.claritas.com/mybestsegments/. Accessed

Jan. 31, 2020.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer,

P., Weiss, R., Dubourg, V., Vanderplas, J., Passos,

A., Cournapeau, D., Brucher, M., Perrot, M., and

Duchesnay, E. (2011). Scikit-learn: Machine learning

in Python. Journal of Machine Learning Research,

12:2825–2830.

Ueno, R., Bryce, R., and Calitoiu, D. (2019). Ranking Clus-

ters of Postal Codes to Improve Recruitment in the

Canadian Armed Forces. In Proceedings of the 18th

IEEE International Conference on Machine Learning

and Application, volume 1, pages 1192 – 1197. Boca

Raton, FL, USA.

Ward, J. H. (1963). Hierarchical grouping to optimize an

objective function. Journal of the American Statistical

Association, 58(301):236–244.

ICORES 2021 - 10th International Conference on Operations Research and Enterprise Systems

314

APPENDIX

Maps of PRIZM Segments

Figure 6: Geographical locations of the postal codes in

PRIZM segment 32.

Figure 7: Geographical locations of the postal codes in

PRIZM segment 37.

Figure 8: Geographical locations of the postal codes in

PRIZM segment 16.

Figure 9: Geographical locations of the postal codes in

PRIZM segment 24.

Figure 10: Geographical locations of the postal codes in

PRIZM segment 63.

PRIZM Segment Descriptions

Below are the descriptions of PRIZM segments de-

ﬁned in the technical documentation provided by En-

vironics Analytics.

PRIZM 32. Mini Van & Vin Rouge represents a col-

lection of younger and middle-aged fami-

lies and couples who live in new exurban

communities beyond Quebec’s big cities.

These households consist of married and

common-law couples. Although most

households are French-speaking, more

than 40 percent are bilingual. Their mixed

educations provide good blue- and white-

collar jobs in the construction and manu-

facturing sectors, as well as management

roles within the public-service sector, re-

sulting in upper-middle-incomes and ac-

tive lifestyles. Residents here have gym

memberships and enjoy sports like hockey,

cross-country and downhill skiing. Af-

ter all that fresh air and exercise, they re-

ward themselves by picking up dinner at

a chicken restaurant or kicking back with

a glass of Shiraz in their single- or semi-

detached homes. For a night out, they

might head to the opera or a popular music

concert; their idea of a vacation is anything

from a resort package to a sightseeing tour

around the U.S. and Caribbean.

PRIZM 37. Younger and middle-aged families com-

prise Trucks & Trades, where skilled

tradespeople and blue-collar workers have

built a comfortable lifestyle while accu-

mulating tidy savings. Concentrated in

Alberta and the Prairies, this segment

has a disproportionate number of oil and

gas workers who have sought out jobs

in resource-rich lands over the past two

decades. What workers may lack in ed-

ucation, they make up for with practical

Identifying Geographical Areas using Machine Learning for Enrolling Women in the Canadian Armed Forces

315

skills in primary industries as well as the

trades and transportation sector. Many

families are younger and middle-aged –

most children are under the age of 15 – and

live in single- and semi-detached houses

built between 1961 and 1990. There’s

also an above-average presence of mobile

dwellings hauled in to accommodate the

sudden inﬂux of workers. When not work-

ing hard, these households play hard: ﬁsh-

ing, hunting, golﬁng, ATVing, snowmobil-

ing and playing baseball, along with other

sports. They also have high rates for own-

ing boats, camping trailers and motorcy-

cles.

PRIZM 16. One of the largest lifestyles in Canada,

Pets & PCs is a haven for younger families

with preschool children in the new sub-

urbs surrounding larger cities. Nearly half

of the children in this segment are under

the age of 10, while many of the maintain-

ers are under 45. Pets & PCs households

have a strong presence of immigrants from

China, the Philippines and India. Few seg-

ments have more new housing than this

group; most residents have settled into a

mix of single-detached, semi-detached and

row house developments. With upscale in-

comes, segment members have crafted an

active, child-centred lifestyle. These fam-

ilies participate in many team sports, in-

cluding baseball, basketball and hockey,

and they shuttle kids and their gear to

games in spacious SUVs–typically newer

models. On weekends, they head to kid-

friendly destinations like zoos, aquariums

and amusement parks. They ﬁll their

homes with an array of computers and

electronic gear, including video game sys-

tems, tablets and just about anything that

will occupy their children while the moms

and dads grab the occasional date night

at the movies or dinner at their favourite

seafood restaurants.

PRIZM 24. Widely dispersed across Canada, Fresh

Air Families is one of the largest

segments–and growing. Found in rapidly

expanding exurban communities, these

neighbourhoods feature a mix of middle-

aged couples and families with children

of a broad spectrum of ages. While

most adults have high school, trade school

or college educations, these dual-income

households enjoy solid, upper-middle-

income lifestyles thanks to positions in

public administration, construction and the

skilled trades. They own single-detached

homes, typically built since 1990, and nine

out of 10 commutes by car to jobs in

nearby suburbs. With its couples and fami-

lies, the segment scores high for a range of

marketplace preferences, frequenting big-

box retailers, large department stores and

discount grocers. Members of Fresh Air

Families enjoy the great outdoors, particu-

larly ﬁshing, boating, canoeing and camp-

ing. Indeed, some of their favourite leisure

activities are evident in their driveways,

typically cluttered with boats, campers or

motorcycles – and pickup trucks to haul

them to parks and campgrounds. But they

also enjoy indoor pursuits like crafting and

knitting.

PRIZM 63. Located in dense, industrial neighbour-

hoods scattered across mid-sized cities,

Lunch at Tim’s consists of singles, fam-

ilies and solo-parent households living in

older single-detached homes, semis and

duplexes. They’re the kind of tight-knit

communities where residents enjoy social-

izing at local eateries like Tim Hortons–

as well as pizza places, burger joints and

ﬁsh-and-chip restaurants. With an unusu-

ally mixed age proﬁle–it’s no longer the bi-

modal segment of the past–Lunch at Tim’s

residents have above-average rates for res-

idents who are single, divorced, separated

or widowed; nearly half the adults in these

neighbourhoods are unattached. Despite

the lower-middle incomes, roughly two-

thirds of households own their homes,

mostly built before 1980. Residents enjoy

quieter pastimes and have high rates for

knitting and woodworking, as well as out-

door activities like hiking and swimming.

When the mood strikes, they might play

a friendly game of curling or splurge on

tickets to a dinner theatre, baseball game

or boat or craft show.

ICORES 2021 - 10th International Conference on Operations Research and Enterprise Systems

316