Is It Possible to Recognize Apple Employees by Their LinkedIn Proﬁle

Picture?

Thanakij Wanavit

and Leslie Klieb

Business and Technology, Webster University Thailand, 1 Empire Tower, Sathorn Rd., Bangkok 10120, Thailand

Keywords:

Convolutional Neural Networks, Emotion Recognition, Employment at Apple, EmoPy, Random Forest

Classiﬁer, Classifying Employment from Emotions.

Abstract:

Samples of images from the portraits on the proﬁles of members of the social media site LinkedIn who live in

the Bay Area of San Francisco were collected and analyzed by the EmoPy package for the presence of seven

emotions. A Random Forest classiﬁer used these probabilities to predict if the members were employed by

Apple or not. Accuracy reached around 62% compared with a naive error rate of 50%. An error analysis shows

that this result is signiﬁcant and robust. A connection between the data and Apple’s organizational culture is

pointed out.

1 INTRODUCTION

Is it possible to ﬁnd out from proﬁle pictures on a so-

cial network where people work? This position pa-

per will suggest that the use of neural networks and

machine learning can indeed accomplish this in some

circumstances. Employment at one of the FAANG

(Facebook, Apple, Amazon, Netﬂix, Google) compa-

nies as a software developer is surrounded by an aura

of mystique. The Internet abounds with people, es-

pecially computer science students, searching for an-

swers to the question of how to get hired by one of

those prestigious companies. Most answers on the

Internet focus on skill sets, ability to answer inter-

view questions, and other cognitive advantages. The

research question of this work tries to answer a re-

lated question focusing on non-cognitive differences.

Is it possible to ﬁnd differences between male soft-

ware engineers working for Apple, Inc. and those

working for other companies in the San Francisco

Bay Area that are sufﬁcient to classify those employ-

ees as Apple or non-Apple employees, and that have

no connection with background, skills, or other obvi-

ous cognitive characteristics? In order to answer that

question, a sample of LinkedIn proﬁles was scraped

from the LinkedIn website searching for current em-

ployment at Apple or not Apple, and for location in

the Bay Area. No other information was collected

https://orcid.org/0000-0001-7291-394X

https://orcid.org/0000-0002-0881-5330

or retained. As an example of a characteristic that is

not cognitive, the research uses machine learning and

neural network technologies to determine a numeri-

cal estimate of the emotional content present in the

photographs. Note that those images are photographs

chosen by the employees themselves with the pur-

pose to present a professional look that is, in com-

bination with the other content in the proﬁle, attrac-

tive for headhunters and others interesting in hiring

developers. The Random Forest statistical algorithm

is trained on a set of those images. It is found that the

software is indeed able to make classiﬁcations with a

small but signiﬁcant success rate.

2 WHY APPLE?

An experiment like this has only a chance to succeed

when a few conditions are met. The core of employ-

ees must be in some sense sufﬁciently distinct from

the employees at other companies. This is in the Bay

Area true for Apple.

Firstly, Apple is sufﬁciently large that a reason-

ably size random sample could be collected. That is

not true for smaller companies.

Secondly, the company is focused on innovation

(Podolny and Hansen, 2020). It renews its products

offerings every one or two years. Steve Jobs estab-

lished for this purpose a culture of competition be-

tween groups inside the company. This is a large dif-

Wanavit, T. and Klieb, L.

Is It Possible to Recognize Apple Employees by Their LinkedIn Proﬁle Picture?.

DOI: 10.5220/0010642300003063

In Proceedings of the 13th International Joint Conference on Computational Intelligence (IJCCI 2021), pages 277-284

ISBN: 978-989-758-534-0; ISSN: 2184-3236

277

ference with other companies with more established

products. It will attract a special kind of developer.

Thirdly, working for the Apple company is glam-

orous and they can pick their staff to a much higher

degree than smaller software companies can do. Dif-

ferently from what might be thought, Apple does not

dominate the software developing industry in Silicon

Valley. There are less developers working at Apple

than at Google in the Bay Area. It was estimated that

in May 2016 there were 41,490 application software

developers and 28,670 systems software developers in

the Bay Area. An estimate for the number of software

developers at Apple is around 12,000-16,000. There-

fore it is possible for Apple to be picky in who they

hire, not only in skills but also as ﬁt to a quite different

company culture. (Anonymous, 2017)

3 PROCESS OVERVIEW

In this short section an overview of the analysis pro-

cess is given. Details and references are in the

Methodology section. The steps are the following:

• Collect LinkedIn proﬁle pictures of three different

samples

– Male Apple employees in the Bay Area of San

Francisco with occupation computer software

(Apple=true, Control=false)

– Male non-Apple employees in the Bay Area of

San Francisco, same occupation (Apple =false,

Control=false)

– Male employees, any occupation, from the

USA ( Apple=false, Control = true)

Those three categories are shown in Table 2.

• The EmoPy package provides in one package ob-

ject detection and recognition of seven emotions:

anger, fear, calm, sadness, happiness, surprise,

and disgust

• Statistical analysis of the differences between the

three samples

4 LITERATURE OVERVIEW

Every company depends for its survival on a steady

stream of new personnel that via a process of on-

boarding will become a good ﬁt in skills, character,

motivation, assimilation of the organizational culture,

and other traits. It is therefore to be expected that

there are differences between employees who work

for a company like Apple and for other companies.

The selection will start already at the reading of the

resumes and looking at LinkedIn proﬁles, and can

lead to differences in the two groups that are ampli-

ﬁed at every stage, from interview and hiring deci-

sion, to turnover and retention (Marsden, 1994) Im-

plicit or explicit biases might be present that inﬂuence

this process (Bendick and Nunes, 2012).

Social media proﬁles have been found to repre-

sent the personality of individuals (Back et al., 2010).

Relationships between characteristics of pictures and

personality were found in Facebook proﬁle pictures

by (Celli et al., 2014). As different companies can

attract different types of people and those differences

can be discernible by AI in the proﬁle pictures on So-

cial Media, it is an interesting issue to see if in the

reverse direction, the proﬁle pictures are sufﬁciently

distinct so that it is possible to classify the employees

according to the company.

5 METHODOLOGY

Sampling was automated using Selenium. It was car-

ried out by searching LinkedIn from an apparent lo-

cation in the USA at the “search for people” tab with

ﬁlter “company: Apple, location: San Francisco Bay

Area, occupation: computer software, sex: male”,

and keeping every second result. Similarly, results

were collected for non-Apple employees in the Bay

Area with the same ﬁlter but not working for Ap-

ple. A control group had as ﬁlter “is in the United

States” and not working for Apple. The search rou-

tine stored the proﬁle picture, occupation, and com-

pany but no other information. There were 851 Apple

employees, 416 non-Apple employees, and 517 “con-

trol” LinkedIn members in the sample.

The freely available open source EmoPy package

(Perez, 2017) from Thoughtworks (Perez, 2018b),

(Perez, 2018a) was the main tool to extract emotions

from the LinkedIn image set. Extraction of the deeper

information in an image is possible using a combina-

tion of a trained object detection Convolutional Neu-

ral Network (CNN) followed by classiﬁcation CNN.

This is demonstrated by an emotion extraction pack-

age such as EmoPy (Perez, 2018c). EmoPy uses ma-

chine learning to recognize the presence and strength

of the expression of seven universal (not depending

on cultural artifacts) standard emotions (anger, fear,

calm, sadness, happiness, surprise, and disgust) in ev-

ery image. The tool itself consists of a number of

CNN layers interspersed with pooling layers that de-

crease the amount of information when it gets too

large. The EmoPy package comes with a series of

default pre-trained models. One of the pre-trained

NCTA 2021 - 13th International Conference on Neural Computation Theory and Applications

278

Table 1: EmoPy parameters.

Layer (type) Output Shape Param

conv2d 1 (Conv2D) (None, 61, 61, 10) 170

conv2d 2 (Conv2D) (None, 58, 58, 10) 1610

maxpooling 2d 1 (None, 58, 29, 5) 0

(Maxpooling2)

conv2d 3 (Conv2D) (None, 55, 26, 10) 810

conv2d 4 (Conv2D) (None, 52, 23, 10) 1610

maxpooling 2d 2 (None, 52, 11, 5) 0

(Maxpooling2)

ﬂatten 1 (Flatten) (None, 2860) 0

dense 1 (Dense) (None, 3) 0

Total parameters: 12,873

Trainable parameters: 12,873

Non-trainable parameters: 0

models available in EmoPy is the Facial Expression

Recognition (FER) model which is a CNN trained

on the FER dataset (Barsoum et al., 2016), which is

based on the FER2013 image set (Perez, nd). A mod-

iﬁed model was made to enable that the model pre-

dicts seven outputs instead of four, a limitation of the

version used here. To ﬁx, the scores of each emotion

of nine models with 2, 3 or 4 emotions were aver-

aged: Model 1: (a)nger, (f)ear, (su)rprise, (c)alm; 2:

(h)appiness, (d)isgust, (su); 3: (a),(f),(su); 4: (a), (f),

(c); 5: (a), (h), (c); 6: (a), (f), (d), 7: (c), (d), (su); 8:

(sa), (d), (su); 9: (a), (h). A new update to the EmoPy

packages seems to make this currently unnecessary.

This was not available at the time of this work. Also a

classiﬁer layer was added. A description of the design

is in (Wanavit, 2021).

Table 1 shows Emopy parameters used in the

runs. All other parameters were kept at default (Pe-

dregosa et al., 2011). Before emotions can be rec-

ognized, the system must ﬁrst separate the face itself

from the surrounding background. A CNN used by

Emopy uses the same concepts as were proposed in

the well-known YOLO Paper (Redmon et al., 2016)

to recognize objects with four methods simultane-

ously, namely pose-robust feature extraction, multi-

view subspace learning, face synthesis based on 2D

methods, and face synthesis based on 3D methods, all

in one fast pass. A CNN like EmoPy can do these

simultaneously together with the emotion recognition

in YOLO mode (combining the face recognition itself

and the emotion recognition) in one pass while using

the Graphics Processor Unit for speed. The Python

code ran in a Jupyter notebook in a virtual machine

provided by Google Colab (Colaboratary, nd). The

data inside Google Colab is automatically backed up

on Google Drive and is available publicly.

Each of the seven emotional features in each

LinkedIn picture is assessed by the CNN and gets a

score which indicates the strength of that emotion.

These scores are normalized so that they sum to 100.

This happens at the end of the pipeline in a ﬂattening

layer. Collectively, the strength of those emotional

features should summarize all the useful information

taken into account from the image in this research

while reducing the amount of data fed into the ﬁnal

layer of image classiﬁer. The last layer is a fully con-

nected layer (Perez, 2018a), and converts the strength

of the found emotions into probabilities, with their

sum per image equal to 100% via a SoftMax activa-

tion function.

so f tmax(z

) =

exp(z

)

∑

j=1

exp(z

)

, (1)

where k is the number of classes, 7 in this case, i=1...k,

and weights and bias have been neglected because

they are difﬁcult to assess here. The SoftMax func-

tion is a generalization into more dimensions of the

sigmoid function from logistic regression that makes

sure that the total probability is smaller than 1. z

the strength of emotion i.

Also, the Keras-Tensorﬂow library of low-level

and high-level API routines was used in this research.

Tensorﬂow also uses the GPU and Colab.

The research photos and temporary data logs of

this work were stored in the S3 object store, a ser-

vice from Amazon Web Service which allows user to

store and distribute ﬁles stored into folder-like buck-

ets. S3 has multiple advanced feature which are ben-

eﬁcial to this type of research including data redun-

dancy, version control, accelerated data transfer, high-

speed web hosting, and permission control. The Dy-

namoDB database service from Amazon was used to

store all data, also the non-structured data. The ser-

vice is low cost and the performance is sufﬁcient for

real time data processing used by the Python code

in this research. The database is highly redundant

with safety measures built-in to prevent accidental

data losses. All forecasts data from EmoPy and other

binary algorithms are stored in DynamoDB within S3

(Amazon, n d), (Sivasubramanian, 2012)

6 ANALYSIS

The output of EmoPy is passed in its last layer

through the SoftMax activation function in order to

convert the scores of the emotion feature set into a

probability. The input for the binary classiﬁer is the

output of the EmoPy neural network which consist of

the probability of each of the seven emotions in each

image. The output of the binary classiﬁer is the pre-

dicted probability of the person working for Apple or

Is It Possible to Recognize Apple Employees by Their LinkedIn Proﬁle Picture?

279

not. This is a number between 0 and 1. 0 means the

person is predicted as certainly not working for Apple

and 1 means the person is predicted with certainty to

work for Apple. If the output is 0.5 or higher, then

this predicts that the person is working for Apple, just

like in binary logistic regression methods. If the result

is correct, then the algorithm gets 1 point for making

the correct prediction and if incorrect, zero points.

The quality of a forecast is usually given by the

confusion matrix (Markham, 2014) and by its accu-

racy. The best performing algorithm was Random

Forest. This is a meta estimator (Breiman, 2001) that

is ﬁtting some decision tree classiﬁers on various sub-

samples of the data set. It uses averaging to improve

the accuracy of the prediction and for control of over-

ﬁtting. For a few other methods see Appendix A.

The Random Forest model was implemented us-

ing source code from scikit-learn (Pedregosa et al.,

2011) available via (Anonymous, ndc).

The analysis was always done by comparing two

sets of data only, Apple versus non-Apple or Apple

versus Control, and not by attempting to classify all

three sets of data simultaneously. Because we were in

charge of the data collection and the choice was made

to do only binary classiﬁcations, it was advantageous

to avoid the accuracy-paradox and balance the sam-

ples so that the base accuracy always was 50%. This

balancing is not always strictly required by all used

algorithms but is recommended (Anonymous, ndb).

The sample set was kept balanced by randomly re-

moving some Apple cases from the 851 collected im-

ages for Apple so that the total number of cases fed to

the model is 416 for the Apple, non-Apple and Con-

trol sets. The group of in total 2 * 416 = 832 Ap-

ple and non-Apple employees was divided randomly

into 666 training cases and 166 test cases (a ratio of

80% versus 20%). This training set of 666 cases was

kept at 50% Apple and 50% non-Apple, so that the

test set was also balanced. An estimate of the error

in the forecast was made by repeating this procedure

1000 times. For each of the 1000 runs, 416 Apple

images were randomly chosen from the whole group

of 851 collected Apple images and then divided ran-

domly into 666 balanced training and 166 balanced

test cases. The model is then trained and made to

make another forecast. Results are plotted and dif-

ferences analyzed using the z-test to see whether the

prediction is signiﬁcantly higher than the 50% base-

line that results from equal-size sample sets and its

improvement over baseline is a reliable estimate.

Figures 1 and 2 show a scatter diagram of the bi-

nary relationships between each pair of emotions for

Control=false. The orange colors are for Apple em-

ployees and the blue colors for non-Apple employees

who are also computer engineers

Figure 1: Correlation of the seven emotions for Apple and

non-Apple. Y-axis (from top to bottom) is surprise, calm,

happiness, sadness, disgust, anger and fear. X-axis is from

left to right surprise, calm, happiness and sadness.

Figure 2: Correlation of the seven emotions for Apple and

non-Apple. Y-axis (from top to bottom) is surprise, calm,

happiness, sadness, disgust, anger and fear. X-axis is from

left to right disgust, anger and fear.

Visual inspection of binary relationships between

all emotions does not show any obviously different

pattern for Apple and non-Apple Control employees.

A negative linear relationship is visible in the plots

between sadness and surprise and between happiness

and anger for both groups. The latter relationship is

intuitively clear, but there is no clear explanation for

the ﬁrst relationship.

Figures 3 and 4 show 3D plots of Apple and non-

Apple employees. There is a suggestion in these plots

that Apple hires more people who score outside the

NCTA 2021 - 13th International Conference on Neural Computation Theory and Applications

280

Table 2: Mean and Standard Deviations Probabilities for all samples.

Apple Control surprise calm happiness sadness disgust anger fear

False False Mean 17.08 20.72 21.23 10.37 1.28 20.29 9.03

SD 6.71 1.67 7.55 10.38 1.00 6.94 2.20

False True Mean 16.67 20.61 20.56 10.98 1.24 20.76 9.19

SD 6.83 1.69 7.80 10.44 1.02 7.26 2.31

True False Mean 17.01 20.73 19.83 10.25 1.29 21.58 9.32

SD 6.86 1.73 7.80 10.33 1.05 7.50 2.36

Figure 3: 3D scatter plot for anger, happiness and disgust,

for Apple and non-Apple employees.

Figure 4: 3D scatter plot for anger, happiness and calm, for

Apple and non-Apple employees.

average. The edges of the 3-dimensional clouds con-

tain mostly Apple employees. One could speculate

that Apple, as a company that depends on innovation

and therefore creativity for its success, is willing to

employ people that are less conformist.

Table 2 gives the mean and standard deviation of

the probabilities for each emotion for all samples. It

is seen that only “calm” has a narrow peak (small

standard deviation relative to the mean) and that dis-

gust and sadness have a very wide shape, with a stan-

dard deviation of the same order of magnitude as the

mean. Differences in averages between Apple and

non-Apple employees are minimal. This makes clas-

sifying difﬁcult. However, the standard deviations

for all emotions are larger for Apple employees than

for non-Apple employees, consistent with the wider

range of emotions that can be seen for Apple em-

ployees in the 3D plots. This supports the idea that

Apple is willing to hire a more diverse range or peo-

ple. This was conﬁrmed by a One-Way ANOVA test

in SPSS between Apple and non-Apple employees.

The p-value for the difference between the means are

p=0.003 for anger and p=0.037 for fear, signiﬁcant at

the 5% level. Both are higher with Apple employees.

Much more interesting, the variances (square of the

standard deviations) are also signiﬁcantly larger, p=

0.005 for anger and p= 0.043 for fear.

One of the 1000 runs yielded, as an example, the

confusion matrix in Table 3. This is a more or less

random, average, example. Its accuracy is very simi-

lar to the most frequent score, as can be seem in Fig-

ure 5. There is no way to ﬁnd the most accurate pre-

diction without trying them all out and getting an idea

of the distribution of the accuracy scores.

Table 3: Confusion Matrix for one of the runs. The accuracy

score is 0.620481.

Ground Truth

Positive Negative

Prediction

Positive 50 33

Negative 30 53

(a) Confusion matrix

precision recall f1-score support

False 0.62 0.60 0.61 83

True 0.62 0.62 0.62 83

(b) Report

166 balanced testing cases, 83 from Apple and 83

from non-Apple software engineers, Random Forest

Is It Possible to Recognize Apple Employees by Their LinkedIn Proﬁle Picture?

281

gave a forecast of 50 true positives, 53 true negatives,

30 false negative, and 33 false positives. For the thou-

sand runs, the lowest accuracy was 51.03%, the high-

est 65.4%, the 25th percentile was at 57.48%, the me-

dian at 59.24%, and the 75th percentile at 61.00%.

The average was 59.18% with a standard deviation of

2.46% and a standard error of σ/

√

n, using a sam-

ple size of 1000, of 0.078. The z-score of the mean

compared with the base rate of 50% is then incredibly

large and the null hypothesis that the Random Forest

method has a 50% base rate impossible for any rea-

sonable level of signiﬁcance. The histogram in Figure

5 shows the Gaussian shape

Figure 5: Accuracy in 1000 runs with the Random Forest

model.

The various emotions contributed in the following

way to the classiﬁcation as in Figure 6.

Figure 6: Contribution of each emotion to the Random For-

est model.

As another metric of the quality of the forecast,

the ROC and AUC are shown for the same run as in

the Confusion Matrix in Figure 7

A comparison of Apple-employees with the con-

trol group gives similar results. There is not much dif-

Figure 7: ROC and AUC for one of the runs.

ference between the Control group and the non-Apple

group. This provides an argument not to handle this

as a classiﬁcation problem between all three groups.

7 DISCUSSION AND

CONCLUSIONS

Only sadness is of less importance in the classiﬁca-

tion of the Random Forest model, further all emo-

tions seem to play a role, see Figure 6. It is tempt-

ing to speculate about the connection of the emotions

in the images and the personal characteristics that can

make a difference in hiring and retention at Apple.

Given that fear and anger seem to make the largest

contributions, consistent with the One-Way ANOVA

test, this points to a connection with Apple’s inter-

nal competition-driven organizational culture. Both

emotions have psychological links with aggression

and perceptions of being attacked or threatened via

the ﬁght or ﬂight (hyperarousal) response. Links be-

tween this response and anxiety and aggression have

been studied extensively by psychologists, especially

in children (Kunimatsu and Marsee, 2012).

Whether this is true or not, this work has shown

that there is a deﬁnite connection between the ex-

pressiveness of emotions in the LinkedIn proﬁle pic-

tures and employment with Apple or others. How-

ever, exploration is made complicated because there

is no simple linear relationship between the probabil-

ity of the emotions as uncorrelated independents and

the probability to work at Apple. If there would have

been, then binary logistic regression analysis would

have worked, but it did not. We don’t know if there are

higher-dimensional “islands” of emotions with cer-

tain probabilities that are connected with a higher or

lower chance to work at Apple, or if the reasons are

NCTA 2021 - 13th International Conference on Neural Computation Theory and Applications

282

even deeper. It is, however, possible that such “is-

lands” reﬂect psychological constructs or biases like

age or race more than a bias towards people showing

certain emotions.

As this is a position paper, we point out that

the data are consistent with an organizational cul-

ture at Apple that is fairly aggressive (anger, fear

and calm are the most discriminating emotions) and

in accordance with ideas from Complexity Science

about companies that are successful in thwarting at-

tacks from their competition and in attacking their

competition. Complexity Science as applied to busi-

ness borrows ideas from the physics of systems far

from equilibrium and from the biology of evolution.

Successful companies operate at the “edge of chaos”,

where interactions with employees and outside con-

tacts are neither too sparse (so that the company is

not cohesive enough) or too tight (which would make

the company ungovernable). This diversity in its em-

ployees and the preponderance of weak links with

people with a different background enable innova-

tion and quick responses to changing circumstances

(Beinhocker, 1997). The wider range of emotions

reﬂect a larger proportion of “not so average” em-

ployees, who in turn are more perceptive of devel-

opments in parts of society and the globe that are

missed in more conventional companies. (Podolny

and Hansen, 2020) describes in detail the type of man-

agement structure needed so that such employees are

still able to inﬂuence the internal debate but also are

independent enough that they can keep functioning as

an antenna for external developments. This is a type

of diversity that goes beyond the inclusion of minori-

ties (Anonymous, nda)

The research focused only on the emotions and

did not collect any other data about the people on the

photographs. In future work we hope to explore if

the enhanced labeling of the FER+ image data set can

improve the classiﬁcation and if the formalism can be

applied to other situations.

ACKNOWLEDGMENTS

The authors want to extend their thanks to Dr. Pin

Pin Tea-makorn for helpful discussions and to the

anonymous reviewers for their suggestions for im-

provement.

REFERENCES

Amazon (n. d.). Amazon s3. https://amazon.com/s3/. In-

formation Page.

Anonymous (August 19, 2017). How many software en-

gineers are currently employed in the Silicon Valley

Bay area? https://askwonder.com/research/software-

engineers-currently-employed-silicon-valley-bay-

area-ran93oji9.

Anonymous (n.da). Inclusion & diversity.

https://www.apple.com/diversity/.

Anonymous (n.db). knn and unbalanced classes.

https://stats.stackexchange.com/questions/341/knn-

and-unbalanced-classes.

Anonymous (n.dc). Random Forest Classiﬁers.

https://scikit-learn.org/stable/modules/generated/

sklearn.ensemble.RandomForestClassiﬁer.html.

Back, M. D., Stopfer, J. M., Vazire, S., Gaddis, S.,

Schmukle, S., Egloff, B., and Gosling, S. (2010).

Facebook Proﬁles Reﬂect Actual Personality, Not

Self-Idealization. Psychological Science, 21(3):372–

374.

Barsoum, E., Zhang, C., Ferrer, C. C., and Zhang, Z. (Oc-

tober 2016). Training deep networks for facial ex-

pression recognition with crowd-sourced label distri-

bution. In Proceedings of the 18th ACM Interna-

tional Conference on Multimodal Interaction, page

279–283.

Beinhocker, E. (1997). Strategy at the edge of chaos. The

McKinsey Quarterly, 1:24–39.

Bendick, M. and Nunes, A. P. (2012). Developing the Re-

search Basis for Controlling Bias in Hiring. Journal

of Social Issues, 68(2):238–262.

Breiman, L. (2001). Random Forests. Machine Learning,

45(3):5–32.

Celli, F., Bruni, E., and Lepri, B. (2014). Automatic Per-

sonality and Interaction Style Recognition from Face-

book Proﬁle Pictures. In MM ’14: Proceedings of the

22nd ACM international conference on Multimedia,

page 1101–1104.

Colaboratary (n.d.). Frequently asked questions.

https://research.google.com/colaboratory/faq.html.

Google Colaboratory.

Kunimatsu, M. M. and Marsee, M. A. (2012). Examining

the presence of anxiety in aggressive individuals: The

illuminating role of ﬁght-or-ﬂight mechanisms. Child

Youth Care Forum.

Markham, K. (March 25, 2014). Simple

guide to confusion matrix terminology.

https://www.dataschool.io/simple-guide-to-

confusion-matrix-terminology/. Terminology.

Marsden, P. (1994). The Hiring Process: Recruit-

ment Methods. American Behavioral Scientist,

37(7):979–991.

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.

Perez, A. (2017). Emopy (0.0.5).

https://github.com/thoughtworksarts/EmoPy. Pack-

age.

Is It Possible to Recognize Apple Employees by Their LinkedIn Proﬁle Picture?

283

Perez, A. (2018a). Emopy: a machine learn-

ing toolkit for emotional expression.

https://www.thoughtworks.com/insights/articles

/recognizing-human-facial-expressions-machine-

learning. Thoughtworks Arts [blog].

Perez, A. (2018b). Recognizing human fa-

cial expressions with machine learning.

https://www.thoughtworks.com/insights/articles/

recognizing-human-facial-expressions-machine-

learning. [blog].

Perez, A. (n.d.). Fer2013 image data set for emopy.

https://github.com/thoughtworksarts/EmoPy/tree

/master/EmoPy/models. Package.

Podolny, J. M. and Hansen, M. T. (2020). How apple is

organized for innovation. Harvard Business Review.

Redmon, J., Divvala, S., Girshic, R., and Farhadi, A.

(2016). You Only Look Once: Uniﬁed, Real-Time

Object Detection. In Proceedings of the IEEE Con-

ference on Computer Vision and Pattern Recognition

(CVPR) , pages 779–788.

Sivasubramanian, S. (May 2012). Amazon dynamodb: a

seamlessly scalable non-relational database service.

In SIGMOD ’12: Proceedings of the 2012 ACM SIG-

MOD International Conference on Management of

Data, page 729–730.

Wanavit, T. (August, 2021). Emopy and classiﬁer diagram.

https://github.com/thanakijwanavit/

emotionExtractionUsingEmopy/tree/

43117fdad2a69c032f5d1b0cff77c2f8a53f7e70.

APPENDIX

Results in the main text were based on the Random

Forest classiﬁer. Here are results for some less well

performing classiﬁer methods. K-Nearest Neighbors

is worst. All results were obtained via the classiﬁer

layer, not with a program like SPSS.

Table 4: Confusion Matrix for XGBoost. The accuracy

score is 0.5722891.

Ground Truth

Positive Negative

Prediction

Positive 50 37

Negative 34 45

(a) Confusion matrix

precision recall f1-score support

False 0.60 0.57 0.58 87

True 0.55 0.57 0.56 79

(b) Report

Table 5: Confusion Matrix for Perceptron. The accuracy

score is 0.51204819.

Ground Truth

Positive Negative

Prediction

Positive 70 17

Negative 64 15

(a) Confusion matrix

precision recall f1-score support

False 0.52 0.80 0.63 87

True 0.47 0.19 0.27 79

(b) Report

Table 6: Confusion Matrix for Support Vector Machines.

The accuracy score is 0.56024096.

Ground Truth

Positive Negative

Prediction

Positive 53 34

Negative 39 40

(a) Confusion matrix

precision recall f1-score support

False 0.58 0.61 0.59 87

True 0.54 0.51 0.52 79

(b) Report

Table 7: Confusion Matrix for K-Nearest Neighbors. The

accuracy score is 0.506024096.

Ground Truth

Positive Negative

Prediction

Positive 39 48

Negative 34 45

(a) Confusion matrix

precision recall f1-score support

False 0.53 0.45 0.49 87

True 0.48 0.57 0.52 79

(b) Report

Table 8: Confusion Matrix for Logistic Regression. The

accuracy score is 0.55421686.

Ground Truth

Positive Negative

Prediction

Positive 49 38

Negative 36 43

(a) Confusion matrix

precision recall f1-score support

False 0.58 0.56 0.57 87

True 0.53 0.54 0.54 79

(b) Report

NCTA 2021 - 13th International Conference on Neural Computation Theory and Applications

284