Digital Discrimination Detection in Ridesharing Services in Rio de

Janeiro City

Raquel T. Paiva, Wendy S. Cataldo, Ana Cristina B. Garcia

and Carlos Eduardo R. de Mello

Centro de Ci

encias Exatas e Tecnologia, Universidade Federal do Estado do Rio de Janeiro, Rio de Janeiro, Brazil

Keywords:

Digital Discrimination Detection, Ridesharing Service, Ontology, Machine Learning.

Abstract:

The urban transport sector has been signiﬁcantly transformed by technology. However, the adoption of these

applications has also brought to light important social issues, including the cancellation of rides due to bias.

The objective of this work is to analyze digital discrimination in light of a complex system and address it

through the analysis of crowd data, which can guide mechanisms to dissuade discrimination in digital ser-

vices. Our main motivation is to answer the following research questions: RQ1: Is there evidence of digital

discrimination in the ridesharing services of Rio de Janeiro city? RQ2: Is it possible to identify the factors

that lead to discrimination? RQ3: What are the key concepts regarding Digital Discrimination detection in a

ridesharing service? RQ4: Could Machine Learning techniques accurately identify discrimination and their

key variables that can be used in actions to mitigate this behavior?

1 INTRODUCTION

Currently, urban mobility has become a major chal-

lenge in large centers. The increase in the number

of cars and public transport that has been suffering

with quality and quantity, that directly inﬂuences ur-

ban optimization with direct consequences on trafﬁc

ﬂow and congestion (Batty, 2012), and impact on the

environment, in addition to having a direct relation-

ship with the restriction of offers of legalized urban

mobility services and guaranteed by public authori-

ties and an increase in irregular transport services.

In addition to these, the population suffers with

individual transport services, such as taxis, which

had high costs and also low quality of service, as

they had a monopoly on this service. Given this

scenario, in 2014 the ridesharing service arrived in

Brazil, where the urban transport sector has been sig-

niﬁcantly transformed by technology, with transport

applications playing a key role in simplifying and ef-

ﬁcient services for passengers (Miroslav Tushev and

Mahmoud, 2020), causing new platforms to emerge,

increasing competition, and allowing the population

to have more transport options at affordable costs.

However, the ridesharing apps only enable the

provision of a service in a peer-to-peer modality, that

https://orcid.org/0000-0002-3797-5157

https://orcid.org/0000-0002-3632-4002

is, between a passenger and a driver, who in turn have

biases, a phenomenon that undermines equality and

accessibility of transport services (Yanbo Ge, 2018).

These biases can lead to a social problem which is

discrimination. This brings us to a question: would

ridesharing app available in Rio de Janeiro city be im-

mune to discrimination related to gender, race, age,

able, class and religious, among other characteristics

(Jorge Mejia, 2020)? Does the ridesharing service

available have other issues that we are not aware of?

To clarify these doubts, we sought answers in crowd-

sourced data and explored comments from users of

this system as a way of understanding the main prob-

lem faced by users.

Then in the present study, we sought any evidence

in the literature related to digital discrimination de-

tection in ridesharing services, where we adopted a

strategy of an exploratory literature review.

The database used was Google Scholar and the

search strings were:

• “discrimination prejudice bias ridesharing appli-

cations science computing”

• ”uber ”gender discrimination”

source:Information source:Systems”

• ”uber ”gender discrimination” source:IEEE”

Filtering for the last 5 years, the search resulted in

1030 articles in the ﬁrst query, 9 articles in the second,

and 6 articles in the last, totaling 1045 articles, which

Paiva, R., Cataldo, W., Garcia, A. and R. de Mello, C.

Digital Discrimination Detection in Ridesharing Services in Rio de Janeiro City.

DOI: 10.5220/0012455800003636

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 1205-1212

ISBN: 978-989-758-680-4; ISSN: 2184-433X

1205

were ﬁltered by analyzing the presence or absence of

keywords in the title and summary, in addition to a

brief analysis of the conclusion, which resulted in the

selection of 14 articles for full reading

From this review, we extract the following re-

search questions:

• RQ1. Is there evidence of digital discrimina-

tion in the ridesharing application used in Rio de

Janeiro city?

• RQ2. It is possible to identify the factors that lead

to discrimination?

• RQ3. What are the key concepts regarding Digital

Discrimination detection in a ridesharing service?

• RQ4. Could Machine Learning techniques accu-

rately identify discrimination and their key vari-

ables that can be used in actions to mitigate this

behavior?

The rest of this paper is organized as follows. Sec-

tion 2 provides a background on discrimination. In

section 3, we describe and present our ontology, fol-

lowed by our methodology in section 4. Next in sec-

tion 5, we present our study results and discussion.

Section 6 addresses the conclusion and limitations of

this study.

2 BACKGROUND

In this section, we will present the main concepts

for understanding the research problem and the tech-

niques and approaches for analyzing the problem in

this work. To analyze the main complaints from users

from one platform of ridesharing service and whether

there is any factor of discrimination related to users

of the service, ﬁrst, it is necessary to understand what

discrimination is, how it occurs, and whether it can

be reﬂected in digital services. By understanding how

and when discrimination can manifest itself and its

provoking “agents”, it is possible to assess whether

this discrimination can be extended to ridesharing

digital services.

In the literature, we ﬁnd two types of discrimina-

tion: direct and statistically or proxy. According

to (Brown, 2019), direct discrimination is carried

out by an “agent” based on observable personal char-

acteristics of the person who suffered prejudice and

causes a negative effect (Murphy, 2002). These per-

sonal characteristics can be race, gender, and sexual

orientation, among others.

Statistically or proxy discrimination can occur

consciously or unconsciously and is carried out when

observable personal characteristics are used to infer

about unobservable measures (Brown, 2019) (John

F. Dovidio, 2000). Also, this kind of discrimination it

is known as a belief-based bias (Monachou and Ash-

lagi, 2019) . For example, when we have a service de-

nied for a young person just because statistically we

know that younger people have a lower income than

people over 30 years old.

Another key concept found in the literature is the

taste-based as one of a potential source of discrimi-

nation. According to (Monachou and Ashlagi, 2019)

taste-based bias occurs when a person is not aware of

his own prejudices and is associated with the absence

of information about a person leading to discrimina-

tion. This form of bias is particularly insidious, as it

operates beneath an individual’s levels of conscious

perception, transforming into a subtle yet powerful

form of discrimination.

Now that we know the main concepts related to

prejudice or bias, it is important to understand what

it is digital discrimination or discrimination in the

online environment occurs when a service is denied

to a person or a group of individuals using his per-

sonal characteristics available on their proﬁle in the

services platform that can be used to identify and dis-

tinguish them such as symbols, colors, images, text,

or graphics (Abramova, 2020).

In the methodology section, we use a term called

Red Line as a classiﬁcation category to represent ar-

eas or neighborhoods that have high rates of violence

or crime.

Machine Learning (ML) is a set of models that

enable systems to learn from a dataset, were they can

be trained with a subset of labeled data, called of su-

pervised learning, or independently where the system

identiﬁes patterns and processes, called unsupervised

learning.

We selected two ML techniques for our experi-

ment, the Naive Bayes (NB), a supervised ML tech-

nique used to classify tasks based on Bayes theo-

rem from statistics, and Support Vector Machine

(SVM), another supervised ML technique used to

classify tasks and data analysis for outlier detection.

In the next section, we adopt an ontology which

is a semantic data structure that captures the relation-

ships and concepts underlying a speciﬁc domain.

3 ONTOLOGY

The present work proposes to understand and ana-

lyze if there is evidence of digital discrimination in

the context of the ridesharing service available in Rio

de Janeiro city and whether this reason is related to

some type of prejudice. In this context, the creation

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

1206

Figure 1: Ridesharing ontology.

of an ontology dedicated to understanding the status

of rides, including the canceling, in order to identify if

the reason was due to prejudice is essential for a more

in-depth and effective analysis of this problem.

Our ontology aims to map the key elements re-

lated to ride status, if there is a discriminatory behav-

ior involved or identiﬁed in the cancellations and the

factors that lead to this behavior, also the characteris-

tics of the actors (drivers or passengers) involved.

In order to better understand our domain we an-

alyze the information about the application service

available on the Internet for driver terms and FAQs

for passengers and drivers.

A user will download the application that was

made available on Android and iOS platforms. The

versions of the two operating systems may vary or

present differences in some features. The user, after

downloading and registering on the application, re-

quests a ride by identifying their origin address (em-

barkation) and entering their destination address (dis-

embarkation) and their payment method and select-

ing the desired fare. After entering this information,

the ridesharing platform will search for the nearest

drivers.

Drivers have a speciﬁc application, where they

the system is made to be in an online mode to receive

rides requests. After starting the system and being

ready, they are able to receive ride requests, with ini-

tial entry information such as the boarding region.

After the ride is accepted by a driver, the platform

shares both passenger and driver information on their

respective applications, for example, the ridesharing

platform provides information to the passenger appli-

cation about the driver who accepted the ride, such as

the name of the driver and information about the ve-

hicle. After that, the ride can be canceled by both. If

there is no cancellation, the passenger is boarded, so

the driver starts the ride in his app until disembarking

the ride, where the status changes to complete, and

then moves on to the billing stage, where depending

on the user’s selected option, it can be done directly

in the app or to the driver in the form of credit, debit

or cash.

After billing, the process moves on to the evalua-

tion stage. This is carried out through a rating system

of 1 to 5 stars, where 1 signiﬁes a poor experience and

5 denotes an excellent one. Additionally, there is an

option to leave a comment, which can be used to ex-

press compliments or report any issues that may have

occurred during the journey. In cases of great dis-

satisfaction with a platform’s service, Brazilian users

tend to adopt a complaints platform called Reclame

Aqui, where users register their complaints and the

platform may or may not provide feedback on the

reported complaints. These comments can vary into

two types, commendation, a positive type, and com-

plaints, a negative type, regarding the provision of the

service which is made up of the driver, his vehicle,

the condition of the service provision and the func-

tioning of the platform application itself. Comments

about the driver can vary about different characteris-

tics, such as their driving mode, their education, and

attitudes, such as rudeness or kindness, to behaviors

that should be banned in society, such as prejudice

and harassment. Comments regarding the vehicle can

be very diverse concerning the vendor, model, and age

of the car to its condition and comfort. Other com-

ments that can be found are about the conditions for

providing the service, including considering the route

the driver took, the trafﬁc encountered, and the oper-

ation of the application itself, such as difﬁculties in

registering crashes, or other difﬁculties in use.

4 METHODOLOGY

In our work, we selected an article that we can use as

a baseline for our study and that can be reproduced

with the data that can represent a ridesharing service

used in Rio de Janeiro city. The research method-

ology adopted by (Miroslav Tushev and Mahmoud,

2020) was to analyze the online feedback from the

actors involved in the ridesharing service (drivers and

passengers). The paper adopted the social network

Twitter to represent this online feedback. Due to a

particular characteristic of the Brazilian population,

our proposal is to use the Reclame Aqui platform to

obtain this feedback online.

Digital Discrimination Detection in Ridesharing Services in Rio de Janeiro City

1207

4.1 Dataset

We extracted 210 complaints from user’s platform on

the Reclame Aqui website through a Python algo-

rithm, using the BeautifulSoap and Selenium Web-

driver libraries. Additionally, we received a total of

750 reviews from a ride-sharing service, where there

were 150 reviews from each rating system from 1 to

5 stars. These anonymous comments were saved in

an Excel spreadsheet locally. This dataset then had a

total of 960 comments, in which pre-processing of the

data was then carried out, such as removing lines that

were brought with the phrase ”Optional Comment”,

where the user did not make any comments in the ap-

plication, just inserted an evaluation on the scoring

system. After removing these lines, the ﬁnal dataset

resulted in 433 comments, where special characters

generated, for example, by keyboard support conﬁg-

ured on the smartphone, were removed.

4.2 Classiﬁcation and Data Analysis

After creating the dataset and pre-processing, we read

and analyzed the 433 comments in pairs, where we

manually classiﬁed each one into categories that we

identiﬁed as shown in Table 1.

Table 1: Classiﬁcation categories.

Aggression; Application;

Harassment; Suitability;

Charge; Cleaning;

Positive comments; Driving mode;

Discrimination; Red line;

Education; Ride status; and

Conservation state; Route.

Of the 433 comments, 163 are multilabel and re-

ceived more than one classiﬁcation, as they contained

complaints from 2 or 3 categories, we separate these

comments totaling a dataset with 630 comments.

Despite the small sample of comments obtained,

it was possible to identify the practice of discrimina-

tion, with a percentage of 1.9 percent, as shown in the

graph in Figure 2. In this classiﬁcation, we consider

discrimination in relation to gender, including LGBT,

ethnicity, ageism, weight, politics, and religion.

Of this percentage, 25 percent of users who suf-

fered discrimination were female passengers, 8.3 per-

cent were elderly passengers, and we also found cases

of harassment of women. All religious and politi-

cal discrimination was practiced by passengers in a

percentage of 41.7 in relation to the total number of

comments found with evidence of discrimination as

Figure 2: Discrimination evidence.

Figure 3: Discrimination radar discrimination.

we can see in Figure 3.

After the analysis of the comments and their man-

ual classiﬁcation, we created a data dictionary with

words and terms associated with each of the cate-

gories. Our next step was to use the Python library

scikit-learn for ML to realize a predictive data analy-

sis for our dataset of comments.

The ﬁrst technique used was the NB model for

classiﬁcation, where we applied the MultiOutputClas-

siﬁer and trained the model with the word dictionary

built from the analysis of comments. The result of

the classiﬁcation presented the following indicators:

F1-score of 0.23227, Precision of 0.3706, Recall of

0.1905, and Accuracy of 0.1905.

We continue our test now adopting the SVM

model, also for a classiﬁcation, where we adopted the

same dictionary of words and obtained the following

indicators: F1-score of 0.14636, Precision of 0.3926,

Recall of 0.1111, and an Accuracy of 0.1111.

The comparison of the results obtained with the

two models adopted is shown in Table 2.

Table 2: Results comparison.

NB SVM

Accuracy 0.1905 0.1111

Precision 0.3706 0.3926

Recall 0.1905 0.1111

F1-score 0.23227 0.14636

These results reﬂect the issues faced with the data

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

1208

obtained to construct our dataset, that is, we obtained

a set of unbalanced and small amounts of data for

the adoption of ML techniques. In order to minimize

this issue, we decided to reduce the number of classes

by grouping them according to the proximity between

them, where the result is shown in Table 3.

Table 3: New categories summarized.

Major category Minor categories added

1. Discrimination Aggression and Harassment

2. Conservation state Cleaning

3. Driver Driving mode, Education,

Suitability, Charge,

and Ride status

4. Application Register and

Driver dissatisfaction

5. Route Red line

6. Positive comments No minor added

After regrouping the categories into a smaller

number, we ran our supervised ML models, NB and

SVM, to reclassify the dataset. We divided our dataset

into a part for training and another for testing, follow-

ing a proportion of 80 percent for training and 20 per-

cent for testing. In order to minimize issues related to

overﬁtting, we used the cross-reference scores func-

tion to ﬁtting the models by computing the scores for

10 times consecutive, as shown in Table 4, in order to

improve the results in the test classiﬁcation, as shown

in Table 5. Also, we shown the confusion matrix and

ROC curves to NB and SVM models, as shown in the

sequence of ﬁgures 4, 5, 6 and 7. As can be seen, the

NB model presented better results compared to the

SVM model.

Table 4: Cross-reference scores results.

NB SVM

Score1 0.81 0.68

Score2 0.76 0.72

Score3 0.69 0.72

Score4 0.67 0.78

Score5 0.72 0.74

Score6 0.69 0.74

Score7 0.70 0.72

Score8 0.73 0.66

Score9 0.66 0.74

Score10 0.70 0.76

Mean 0.71 0.72

The problem of class imbalance is a common

challenge in the task of identifying discrimination in

comments, as it can affect the quality of the model.

To solve this problem, we apply the RandomUnder-

Sampler technique from the imbalanced-learn library

Table 5: Comparison of supervised model test results.

NB SVM

Accuracy 0.6385 0.47694

Precision 0.6304 0.4622

Recall 0.6385 0.4769

F1-score 0.6274 0.4645

to balance the model training. We then build an

SVM classiﬁer with speciﬁc class weights ’Applica-

tion’: 5, ’Conservation state’: 7, ’Discrimination’:

10, ’Driver’: 1, ’Positive comments’: 25, ’Route’: 7.

Class weights are assigned so that minority classes

have a greater weight, which contributes to increas-

ing model performance in these classes. With these

adjustments, we created a balanced and adaptive strat-

egy to deal with class imbalance in our dataset. These

changes improved the model’s performance in identi-

fying discrimination in the comments that formed our

dataset, offering a more robust and efﬁcient solution

to tackle the unbalance problem. Therefore, after ad-

justing the speciﬁc class weights, our results changed

to an F1-score of 0.4645 to 0.6681. Table 6 shows the

complete new results.

Table 6: Comparison of supervised model test results.

NB SVM SVM (adjusted)

Accuracy 0.6385 0.4769 0.6846

Precision 0.6304 0.4622 0.6594

Recall 0.6385 0.4769 0.6846

F1-score 0.6274 0.4645 0.6681

Figure 4: Confusion matrix NB model results.

Digital Discrimination Detection in Ridesharing Services in Rio de Janeiro City

1209

Figure 5: Confusion matrix SVM model results.

Figure 6: ROC curve for NB model.

5 STUDY RESULTS AND

DISCUSSION

Digital discrimination in shared services has been ad-

dressed from different aspects, but the main one found

in the literature was diagnosis as in (Abramova, 2020)

(Brown, 2019) (Yanbo Ge, 2018). But approaches

were also found from the aspect of the information

system where the interest was divided into identifying

biases in the algorithms adopted by shared ride plat-

forms as in (Pandey and Caliskan, 2021) and identify-

ing biases and discrimination manifested by applica-

tion drivers as in (Jorge Mejia, 2020), by passengers

as in (Alex Rosenblat, 2017) and by biases expressed

by both as in (Miroslav Tushev and Mahmoud, 2020).

Shared ride service platforms, with the aim of re-

ducing discrimination on the part of drivers, began

Figure 7: ROC curve for SVM model.

sending the least amount of information about the

passenger to drivers when distributing the ride re-

quest. However, it was still possible to observe dis-

crimination shortly after the acceptance and sharing

of passenger characteristics such as name, gender,

score, origin, and destination addresses (Jorge Mejia,

2020).

Another point addressed in (Pandey and Caliskan,

2021) was the bias in the algorithm of shared ride

systems, where the price of ride fares varies not only

with demand, but also with their location of origin or

destination, where positive relationships were found,

that is, it was identiﬁed that locations with a higher

rate of acceptance of rides had higher fares, but also

negative relationships, where locations with a higher

rate of white population have lower rates while loca-

tions with a higher rate of non-white population have

higher rates. In (Jorge Mejia, 2020), the fare value is

identiﬁed as one of the points to reduce the cancel-

lation rate due to discrimination. The study suggests

making this cost explicit as an attempt to reduce bi-

ased behavior. An interesting mitigation action found

in the literature is a ride distribution model based on

learning the history of acceptance and cancellation,

that is, promoting a pairing of passengers and drivers,

not only based on a scoring system but also with the

identiﬁcation of a bias in these platform users and

with this it would be possible to expose them, accord-

ing to (Monachou and Ashlagi, 2019).

Comparing the results found in the literature with

the data and the ontology analyzed in our study, we

suggest some points for discussion, as shown in Fig-

ure 6. The ride-sharing services available in the city

of Rio de Janeiro present several differences from

platform to platform. Based on our study, 7.4 per-

cent of the complaints found were in relation to dis-

counts not being applied or rides canceled due to the

choice of a discounted ride. This difference suggests

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

1210

that users have different perceptions if this mitigating

action works as expected or the dynamic system ap-

plied by this speciﬁc platform service, as we found in

(Jorge Mejia, 2020) and (Pandey and Caliskan, 2021).

Figure 8: Ridesharing discrimination detection discussion.

Another interesting point, which at the same time

corroborates the analysis with the studies found in the

literature, was in relation to the provision of passen-

ger information, such as origin and destination ad-

dresses, to the driver only after accepting the ride as

(Miroslav Tushev and Mahmoud, 2021), (Yanbo Ge,

2018), (Brown, 2019) and (Abramova, 2020). How-

ever, this was a point where we found divergent opin-

ions from the user’s platform. Passenger information

is made available after acceptance of the ride as a way

to mitigate discrimination, however, discrimination is

possible to identify that cancellation still occurs after

this information is made available, and when not, pas-

sengers report that the service provided is impacted,

causing embarrassment, discomfort, and insecurity to

the passenger who is disembarked outside the location

requested in the application.

Also, two indicators suggest a more in-depth anal-

ysis, as it was not possible to identify whether there

was direct or statistical discrimination by class or eth-

nicity. The largest of them, with 20.8 percent of com-

plaints, were related to charging, where they were as-

sociated with passenger complaints regarding drivers

who canceled the ride or did not want to use the dis-

count selected by the passenger when requesting the

ride. Another index that it was not possible to de-

termine direct or statistical discrimination by class or

ethnicity for the 1.3 percent of complaints categorized

as Red Line, that is, where the destination address is

located in communities or their surroundings. This

indicator may be more associated with public safety

issues but also it can hide discrimination behavior.

Furthermore, we were able to observe that the NB

model was the best compared to the SVM, until we

adjusted the class weights to solve the unbalanced

class problem. Additionally, we were unable to use

SVM to identify outliers in our data or obtain better

results due to the size of the datasets with both mod-

els.

6 CONCLUSIONS

In this study, it was possible to analyze that the main

problem of this research is a topic of great relevance

to society and there are opportunities to address it in

the information system in order to promote mecha-

nisms that reduce discrimination of any type, be it

racial, gender, sexual orientation, religious or politi-

cal association, of way to eradicate this behavior that

is harmful to society. Our study, combined with an ex-

ploratory analysis of the state of the art in literature,

proposed to answer the following questions:

• RQ1. Is there evidence of digital discrimination in

the ridesharing application used in Rio de Janeiro

city? Based on our analysis, it was possible to

conclude that yes, there is evidence of digital dis-

crimination in the ridesharing services of the city.

• RQ2. Is it possible to identify the factors that lead

to discrimination? Yes, it was possible to iden-

tify that there are factors associated with preju-

dice in particular towards women, with the com-

ments, it was possible to identify that the majority

of drivers are men, we found only 4 comments

with reference to a driver woman, representing

0.63 percent, and 50 percent with positive com-

ments.

• RQ3. What are the key concepts regarding Digital

Discrimination detection in a ridesharing service?

These concepts were identiﬁed in our analysis of

the domain, where we proposed an ontology about

it.

• RQ4. Could Machine Learning techniques accu-

rately identify discrimination and its main vari-

ables that can be used in actions to mitigate this

behavior? Yes, it is possible to use ML models to

accurately identify discrimination in ride-sharing

services. We were able to observe that, due to the

size of our data set, with just a small adjustment to

reduce the number of categories used for classiﬁ-

cation, we already improved the results presented

by both models. If it is possible to increase the

size of the data set we can expect these results to

improve even further, in addition, if we get a data

set large enough to apply an unsupervised learn-

ing model, it will be possible to compare the re-

sults between the supervised and unsupervised, in

addition to analyzing the identiﬁed patterns and

Digital Discrimination Detection in Ridesharing Services in Rio de Janeiro City

1211

behaviors and checking whether or not it is pos-

sible to identify outliers in our data and thus key

variables for our problem.

As (Miroslav Tushev and Mahmoud, 2020), one

limitation faced was the amount of data, plus the ab-

sence of user information for analysis, as all com-

ments do not contain information and personal char-

acteristics, it does not allow for a more in-depth anal-

ysis of some indicators that may or may not be related

to discrimination due to prejudice, but it was not evi-

dent.

For future work, we propose to enlarge our dataset

to include the complaints from Reclame Aqui web-

site for more ridesharing services offered in Brazil

and from other sources and then we can compare ML

supervised versus unsupervised models for classiﬁca-

tion purposes and to identify outliers in our data anal-

ysis. Also, we can explore the dataset to evaluate

and compare results on the detection of discrimina-

tion between different cities in Brazil, and if the dif-

ferences between ridesharing platforms can increase

or decrease the practice of discrimination.

REFERENCES

Abramova, O. (2020). No matter what the name, we’re all

the same? Electronic Markets.

Alex Rosenblat, e. a. (2017). Discriminating tastes: Uber’s

customer ratings as vehicles for workplace discrimi-

nation. In Policy and Internet.

An Yan, B. H. (2019). Fairst: Equitable spatial and temporal

demand prediction for new mobility systems. In ACM

SIGSPATIAL International Conference on Advances

in Geographic Information Systems.

An Yan, B. H. (2020). Fairness in practice: A survey on

equity in urban mobility. In A Quarterly bulletin of the

Computer Society of the IEEE Technical Committee

on Data Engineering.

Batty, M. e. a. (2012). Smart cities of the future. In The

European Physical Journal.

Brown, A. E. (2019). Prevalence and mechanisms of dis-

crimination: Evidence from the ride-hail and taxi in-

dustries. In Journal of Planning Education and Re-

search.

John F. Dovidio, e. a. (2000). Reducing contemporary prej-

udice: Combating explicit and implicit bias at the in-

dividual and intergroup level.

Jorge Mejia, C. P. (2020). When transparency fails: Bias

and ﬁnancial incentives in ridesharing platforms.

Miroslav Tushev, F. E. and Mahmoud, A. (2020). Digital

discrimination in sharing economy. In International

Requirements Engineering Conference (RE).

Miroslav Tushev, F. E. and Mahmoud, A. (2021). A sys-

tematic literature review of anti-discrimination design

strategies in the digital sharing economy. In IEEE

Transactions on Software Engineering.

Monachou, F. G. and Ashlagi, I. (2019). Discrimination in

online markets: Effects of social bias on learning from

reviews and policy design. In International Confer-

ence on Neural Information Processing Systems.

Murphy, S. A. (2002). Appendix b: Audit studies and

the assessment of discrimination. National Research

Council.

Pandey, A. and Caliskan, A. (2021). Disparate impact of ar-

tiﬁcial intelligence bias in ridehailing. In Conference

on AI, Ethics, and Society.

Yanbo Ge, e. a. (2018). Racial discrimination in transporta-

tion network companies. In Journal of Public Eco-

nomics.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

1212