Jorge Iván Pincay-Ponce
Universidad Laica Eloy Alfaro de Manabí, ULEAM
Manta, Manabí, Ecuador
María Roxana Martínez
Universidad Abierta Interamericana, UAI
CABA, Buenos Aires, Argentina
Wilian Richart Delgado-Muentes
Universidad Laica Eloy Alfaro de Manabí, ULEAM
Manta, Manabí, Ecuador
Juan Alberto Figueroa-Suárez
Universidad Laica Eloy Alfaro de Manabí, ULEAM
Manta, Manabí, Ecuador
DOI: https://doi.org/10.56124/encriptar.v7i14.009
ABSTRACT
This
paper aims to redesign the analysis of the “Speed Dating” dataset, which was
part of the research titled “Gender Differences in Mate Selection: Evidence
from a Speed Dating Experiment,” presented by Raymond Fisman, Sheena Iyengar,
Emir Kamenica, and Itamar Simonson in The Quarterly
Journal of Economics, the oldest professional journal of economics in the
English language, in 2006. Based on the theory of "perceived
availability," which suggests that people are more likely to find those
who seem more attainable or interested in them to be attractive, logistic
regression and the CatBoost ensemble method were employed to uncover patterns
that appear influential in the decisions of individuals of the opposite sex
regarding the potential for a future relationship from a four-minute speed
dating social experiment. The findings indicate that, in general, individuals
prioritize the following in their potential partners, from most to least
important: attractiveness, perceived compatibility, shared interests, sense of
humor, ambition, satisfaction with acquaintances (indicative of sociability),
TV interests, sincerity, and partner's age. These results report an accuracy of
over 80% with Logistic Regression and 88% with the CatBoost ensemble method.
The tool used in model development was Orange Data Mining 3.37.
Keywords: Matchmaking, ensemble, speed
dating
CatBoost y Regresión Logística como enfoques de aprendizaje automático
en el Matchmaking y la Disponibilidad Percibida
Resumen
El presente trabajo
tiene como objetivo rediseñar el análisis del conjunto de datos “Speed Dating”,
que fue parte de la investigación titulada “Gender Differences in Mate
Selection: Evidence from a Speed Dating Experiment” presentada por Raymond
Fisman, Sheena Iyengar, Emir Kamenica y Itamar Simonson, en el The Quarterly
Journal of Economics, la más antigua revista de economía en idioma inglés, en
2006. Con base en la teoría de la "disponibilidad percibida" que
indica que las personas son más propensas a considerar atractivas a aquellas
que parecen más alcanzables o interesadas en ellas, se ha empleado regresiones
logísticas y el método de ensemble CatBoost, para generar con ellos un trabajo
conjunto y descubrir patrones aparentemente influyentes en la decisión de
personas del sexo opuesto sobre llevar una eventual relación de pareja a partir
de un experimento social de citas rápidas de cuatro minutos. Se encontró que en
general, los compañeros y compañeras de las parejas privilegian en el orden de
más a menos lo siguiente en sus parejas: atractivo, probabilidad de
compatibilidad, intereses comunes, divertido o divertida, ambición,
satisfacción con conocidos (indicativo de asocialidad), intereses de TV,
sinceridad, edad de la pareja. Estos resultados reportan una exactitud superior
al 80% con Regresión Logística y 88% con el método de ensamble CatBoost. La
herramienta utilizada en la elaboración del modelo fue Orange Data Mining 3.37.
Palabras clave: Emparejamiento,
métodos de ensemble, citas rápidas.
1. Introduction
The beginning of a relationship involves moving
from a state of indecision to deciding whether to establish a relationship.
Although this is a topic relevant to everyone, there has been relatively little
research on these initial moments (McFarland et al., 2024), particularly
involving machine learning algorithms. However, research from the 1990s
indicated that men placed more emphasis on physical attractiveness rather than
intelligence or ambition, while women placed greater emphasis on income
potential, considering attributes such as ambition, intelligence, and social
status, especially when choosing a partner for long-term relationships. (Buss & Schmitt, 1993; Regan, 1998).
This work aims to redesign the analysis of the
'speed dating' dataset, which was part of the research titled 'Gender
Differences in Mate Selection: Evidence from a Speed Dating Experiment' by
Raymond Fisman, Sheena Iyengar, Emir Kamenica, and
Itamar Simonson, published in The Quarterly Journal of Economics, the oldest
professional economics journal in the English language, in 2006.
Today, the dataset used by the researchers,
which initially applied linear regression models and occasionally separated the
analysis between the 4,194 men and 4,184 women who participated in the study,
is available from various websites.
Our redesign of the original analysis, at the
data preprocessing level, involves generating new categorical attributes to
utilize columns with over 40% missing values and others where numerical values
were used but lacked numerical significance, serving as textual response
options instead.
Although the importance of class balancing in
classification problems is recognized (Fernández et al., 2018), it is not employed
in this research. To minimize noise in the data, the approach focuses on
preventing overfitting or underfitting through LASSO regularization in the
logistic regression algorithm and RIDGE regularization in the CatBoost ensemble
algorithm, along with other hyperparameters.
In the original paper, the researchers used the
attribute 'Decision to pursue or not pursue a relationship' as the class
variable. Specifically, they analyzed the data using linear regression models,
arranged 22 in-person speed dating sessions of four minutes each, and
quantified each participant’s expectations regarding how many people they
thought would be interested in dating them. (Fisman et al., 2006).
It is believed that these inputs have been
foundational for popular online platforms that help people find suitable
partners, using recommendation systems designed to assist users in discovering
other individuals who might also be interested in them. (Kleinerman et al., 2018).
This research has chosen to use the classic
'speed dating' dataset due to the advantage of having social profile data of
users who were physically observed. It is assumed that certain forms of
communication reveal a person’s relational state, whether they are indecisive,
desire a relationship, or do not want one. Additionally, certain forms of
communication can persuade individuals to transition between relational states,
overcoming indecision and reaching a definitive relational decision. (McFarland et al., 2024).
Previous work proposing reciprocal
recommendation systems for online dating websites is highly valued. Based on
the theory of 'perceived availability,' it is believed that people are more
likely to find attractive those who seem more attainable or interested in them.
(Association for the
Advancement of Artificial Intelligence, 2018; Brannan & Mohr, 2018; Sharabi
& Dorrance-Hall, 2024; Ye et al., 2024; Zheng et al., 2022),
Reciprocity is a crucial factor in relationship
formation. Expectations and behaviors in dating have evolved, with a growing
focus on mutual understanding and emotional compatibility. This encourages the
study of how mutual interest develops in these interactions. (Brannan & Mohr,
2018). This is why this
research focuses on the perception of each participant’s potential partner in
deciding a possible match. Specifically, the target column to be analyzed is
the partner’s early decision on the day of the event.
Today, it is known that Machine Learning can
identify patterns in data that are not apparent to humans. In the context of
dating, it can detect subtle compatibilities between users based on less
obvious behaviors and characteristics. The use of Machine Learning for
identifying potential partners is known as matchmaking. (Hayashi et al., 2023; McFarland et al., 2024).
This paper uses the original dataset and
presents a computational approach based on Logistic Regression with Lasso
Regularization. This approach is chosen due to the model’s high
interpretability, the implicit feature selection provided by the
regularization, and the consequent clear communication of the results. (Weigard & Spencer, 2023).
Additionally, since many features in the dataset
were discretized, the CatBoost classifier, based on ensemble methods, is used.
CatBoost is valued for its robustness against overfitting and its ability to
provide competitive performance compared to other boosting algorithms like
XGBoost and LightGBM. Its innovations in handling categorical variables
significantly reduce the need for traditional techniques such as one-hot
encoding, which can increase problem dimensionality and training time. (Joshi
et al., 2021; Pincay Ponce et al., 2024; Prokhorenkova et al.,
2019).
2. Methodology
2.1 Data Preparation
This section documents the changes made in the
data preparation beyond those performed on the original “speed dating” dataset (Fisman et al., 2006):
§ Discretization of
Numerical Features: Numerical features were discretized due to the
use of linear regression models to address the problem (Fisman et al., 2006). High numbers did
not imply better or worse but merely different options, which could introduce
noise into the results (Pincay-Ponce et al., 2020). Features
discretized include gender, race, their goal on a date, frequency of dates,
frequency of outings, field of study, the quality-price ratio of the school
they graduated from, average family income, and whether it matters if their
partner is of the same race, among others.
§ Omission of 1-10
Rating Features: Features rated on a 1-10 scale, where 1 is terrible
and 10 is excellent, were omitted. These ratings reflect what each person
believes is most important to members of the opposite sex when deciding to date
someone, concerning six key characteristics: attractiveness, sincerity,
intelligence, charm, ambition, and shared interests. The feature concerning the
number of past dates was also omitted because 92% of participants did not
provide this information.
§ Retention
and Omission of 1-10 Scale Features: Features on a
1-10 scale were retained, while equivalent scales in 1-100 were omitted.
Retained features include beliefs about how their partner (regardless of
gender) perceives them, how they think men and women perceive the six
attributes, how they think their date perceives them, their self-perception,
their rating of a yes or no during the experimental date, satisfaction with
known people, and their self-perception at the beginning of the date.
§ Scaling
of Perception Features: Features related to the perceived qualities of what
their partner might have, and the general perceptions of other men and women
were scaled to integer ranges of 1-10, as they were originally in 1-100. This
allows discretization into only ten possible values.
§ Scaling
of Match Expectation Features: Features regarding match
expectations were scaled to integer ranges of 1-10, from an original range of
1-20, thus discretizing them into only ten possible values.
§ Imputation
of Missing Data: Columns with missing data that were not omitted from
the analysis were imputed with the median for numerical data or the mode for
categorical data, as these imputation options are less sensitive to outliers (Pincay
Ponce et al., 2024).
§ Creation
of Categorical Features: A categorical feature was created
to identify the initial correlation of interests in a potential partner,
categorized as highly similar (correlation greater than 0.6), highly different
(correlation less than 0.6), or of low relevance in other cases.
§ Creation
of Field of Study Correspondence Feature: A feature was
created to identify the match between a person’s field of study and their
desired career path.
2.1 Regarding data modeling with logistic
regression
The models, including data preprocessing, were
developed in Orange Datamining software. 8378 instances (100%) and 97 features
of 187 originally available were considered for analysis, even so the resulting
data set was of high dimensionality, so regularized Logistic Regression with
Least Absolute Shrinkage and Selection Operator (LASSO, L1) was used, thus
generating a selection of features by setting Odds to zero (Pincay Ponce, 2023), and, consequently,
setting to zero the possibility that certain pairs of characteristic-value
combinations affect the YES/NO decision to make an appointment on the day of
the event (Pincay Ponce, 2023).
The logistic regression is modeled as
Where w represents the vector of coefficients or
weights associated with each of the features
Where
Table 1. Logistic
regressions, regularization hyperparameters and percentage of accuracy
achieved.
|
Subject of analysis |
Regularization |
|
Accuracy % |
|
All participants |
L1, LASSO |
0.003 |
81.6 |
|
Female peers |
L1, LASSO |
0.007 |
85.0 |
|
Male peers |
L1, LASSO |
0.006 |
80.0 |
Source: Investigation
With regularization strength values
Now, the Odds Ratio,
Where -z is the negative of each log-odd of the
coefficients and e is Euler's irrational numerical constant (2.71828).
2.2 Regarding data modeling with CatBoost
CatBoost is a gradient boosting algorithm that is particularly
effective at handling data sets with categorical features, as is the case for
most features after they have been preprocessed for this research. (Pincay Ponce et al., 2024; Prokhorenkova
et al., 2019). Furthermore, it differs from other gradient boosting
such as XGBoost and LightGBM because it builds balanced trees that are
symmetric in structure, this means that at each step, the same split feature
pair that results in the lowest loss is chosen and applied to all nodes at that
level. The hyperparameters considered are:
Table 2. CatBoost and
regularization hyperparameters.
|
Hyperparameter |
Value |
Role in the Algorithm |
|
Number
of Trees, NNN |
100 |
Controls the model's complexity and its capacity to fit the data. |
|
Learning Rate Determines the size at which the weights of the n trees are updated. |
0.2 |
0.2 is
a high learning rate, which can accelerate training but may increase the risk
of overfitting; however, in this case, overfitting was not present. |
|
Reproducible
Training |
True |
Ensures that the training results are reproducible. |
|
Regularization
L2, RIDGE. |
0.19 |
0.19 is
a low value that does not address overfitting but rather underfitting, which
is relevant to this research. |
|
Maximum
Tree Depth |
5 |
5 is a moderate value that balances accuracy and manages overfitting
or underfitting. |
|
Feature Subset |
0.5 |
A
medium value (50%) that reduces feature correlation and improves feature
eligibility. |
Source: Investigation
In the context of CatBoost, Ridge
regularization
Where
To facilitate the interpretation of
CatBoost, SHAP values, which stands for Shapley Additive Explanations, were
used. These address the issue that models often produce output values that are
not easily interpretable because they focus on the 'how much?' of the problem,
meaning the reasons behind the outputs are unknown. SHAP values illustrate how
each feature affects the prediction, even for complex methods such as gradient
boosting (as with CatBoost) or neural networks.(Lundberg, 2018; Van
den Broeck et al., 2022).
3. Resulted
Based on
the feature selection provided by LASSO regularization in Logistic Regression,
we present the probabilities, from highest to lowest, of the influence of
feature-value pairs on the decision on the first day of the event by female
partners and male partners, regarding whether they will eventually match, i.e.,
that first impression. Table 3 shows the alignment in preferences between the
partner's view of the participant's interests and what they consider
attractive. Female partners valued their interest in concerts, TV, music,
sports, and the age of both parties. Male partners valued the age of the women,
but not their own
Table 3. Influence of features – value on the
probability that partners decide to match.
|
|
Women's companions |
P.% |
Companions of men |
P.% |
|
1 |
RatingByPartnerLikeYou |
0.61 |
RatingByPartnerLikeYou |
0.57 |
|
2 |
RatingByPartnerAttractive |
0.59 |
RatingByPartnerAttractive |
0.55 |
|
3 |
RatingByPartnerProbalityMatch |
0.53 |
RatingByPartnerSahredInterest |
0.52 |
|
4 |
RatingByPartnerSahredInterest |
0.50 |
RatingByPartnerFun |
0.52 |
|
5 |
YourInterestingConcerts |
0.50 |
RatingByPartnerProbalityMatch |
0.51 |
|
6 |
YourInterestingTv |
0.50 |
YourYesNoDuringDatingAttractive |
0.51 |
|
7 |
YourInterestingMusic |
0.50 |
ProbableThatPersonAccept |
0.50 |
|
8 |
YourInterestingTVsports |
0.50 |
Self-perceptionPartnerFunTime1 |
0.50 |
|
9 |
Self-perceptionPartnerSincereTime1 |
0.50 |
Self-perceptionPartnerIntelligentTime1 |
0.50 |
|
10 |
Self-perceptionPartnerSharedInterestTime1 |
0.50 |
YourAge |
0.50 |
|
11 |
Self-perceptionPartnerAmbitiousTime1 |
0.50 |
Self-perceptionPartnerSincereTime1 |
0.49 |
|
12 |
Self-perceptionPartnerIntelligentTime1 |
0.49 |
Self-perceptionPartnerAmbitiousTime1 |
0.49 |
|
13 |
AgePartner |
0.49 |
Self-perceptionPartnerAttractiveTime1 |
0.49 |
|
14 |
YourAge |
0.49 |
AttractivePartner |
0.46 |
|
15 |
Self-perceptionPartnerAttractiveTime1 |
0.49 |
match=No |
0.13 |
|
16 |
Self-perceptionPartnerFunTime1 |
0.49 |
|
|
|
17 |
YourInterestingTheater |
0.49 |
|
|
|
18 |
AttractivePartner |
0.48 |
|
|
|
19 |
RatingByPartnerSincere |
0.48 |
|
|
|
20 |
match=No |
0.29 |
|
|
We present the probabilities, from highest
to lowest, of the influence of feature-value pairs on the partners, without
distinction of gender.
Table 4. Influence of features – value on the overall
probability that partners decide to match.
|
|
All |
P.% |
|
1 |
RatingByPartnerAttractive |
0.57 |
|
2 |
RatingByPartnerProbalityMatch |
0.52 |
|
3 |
RatingByPartnerSahredInterest |
0.51 |
|
4 |
RatingByPartnerFun |
0.50 |
|
5 |
RatingByPartnerAmbitious |
0.50 |
|
6 |
SatisfactionWithAcquaintances |
0.50 |
|
7 |
YourInterestingTVsports |
0.50 |
|
8 |
RatingByPartnerSincere |
0.50 |
|
9 |
AgePartner |
0.50 |
|
10 |
Self-perceptionPartnerSincereTime1 |
0.50 |
|
11 |
Self-perceptionPartnerIntelligentTime1 |
0.50 |
|
12 |
Self-perceptionPartnerFunTime1 |
0.49 |
|
13 |
Self-perceptionPartnerAttractiveTime1 |
0.49 |
|
14 |
YourAge |
0.49 |
|
15 |
Self-perceptionPartnerAmbitiousTime1 |
0.49 |
|
16 |
AttractivePartner |
0.47 |
|
17 |
match=No |
0.21 |
The confusion matrix in
Figure 1 shows the accuracy results for each class. The most used metric
in a classification problem like this is accuracy. For illustration, TP stands
for true positive, referring to the set of instances for which the model's
prediction was correct. True negative (TN) refers to the set of instances for
which the prediction was correct; false positive (FP) is the number of
instances where the prediction was incorrect, and false negative (FN) is the
number of instances where the prediction was incorrect.
Therefore, accuracy is
calculated as
Figure 1. Accuracy results for each class according to the
CatBoost algorithm.
Table 5. Accuracy,
precision, recall, and obtained times reported by all models. The period is
used as the decimal separator.
The interpretation of CatBoost has benefited from the
SHAP framework proposed in 2016 by Scott Lundberg and Su-In Lee (2017), because SHAP proved to be an easy and theoretically
sound way to understand the predictions of any model. Figure 2 illustrates
these details, but it does not specify the feature-value pairs, which can often
be inferred from the logistic regression results.
Figure
2. Influence of features on the classification performed
by the CatBoost algorithm.
4. Conclusions
There is still much to learn about attraction
and relationship formation, but the innovative contribution provided by machine
learning is valuable for exploring dyadic behavior, which can offer new
empirical insights into how people choose partners and form relationships. We
have introduced a foundation for generating new research questions on first
impression formation, romantic rivalries, and affiliative behaviors. Here, we
highlight the use of data, such as colleges with a high Quality-Price ratio for
tuition, which is valued numerically between 800 and 1600—the best value—as it
provides an indicator of each participant's economic expectations.
Following this, in general, individuals
prioritize the following traits in their partners, from most to least
important: attractiveness, perceived compatibility, shared interests, sense of
humor, ambition, satisfaction with acquaintances (indicative of sociability),
TV interests, sincerity, and the partner's age. Additionally, the first
impression that the other person projects in terms of sincerity, intelligence,
grace, and attractiveness is also considered. Subtle gender differences, which
were presented in the results section, do exist.
It is acknowledged, but not studied in this
research, that on dating platforms, empirical evidence suggests that the
likelihood of a user being recommended by the platform's algorithm increases
significantly, and perhaps biasedly, with the user's popularity. Therefore, the
analysis presented by us is more traditional or conservative.
It is crucial to acknowledge the limitations of
this study and its specific context. The data originate from a speed dating
experiment, which may not fully reflect partner selection processes in other
environments. Furthermore, technological advancements and social changes since
the original data collection may have influenced dating preferences and
behaviors. Future studies should consider replicating this analysis with more
recent and diverse datasets, as well as exploring how algorithms in online dating
platforms influence matching decisions. Additionally, investigating how
preferences and behaviors vary across different cultures and demographic groups
would be valuable. These considerations are particularly relevant in the
context of machine learning, as the quality and representativeness of training
data significantly impact model performance and generalizability. Moreover, the
dynamic nature of human behavior and societal norms underscores the importance
of continually updating our models and methodologies to ensure they remain
relevant and accurate in capturing the complexities of human mate selection
processes.
5. Future research
Future research directions should encompass a
multifaceted approach to understanding the complexities of mate selection in
the modern era. Firstly, exploring how perceptions of availability evolve over
time in long-term relationships could provide valuable insights into
relationship dynamics. Secondly, investigating the impact of social media and
dating applications on partner selection strategies is crucial in our
increasingly digital world. Thirdly, analyzing how cultural factors influence
the prioritization of attributes in mate selection would offer a more
comprehensive, global perspective on this phenomenon. Fourthly, studying the
influence of artificial intelligence and recommendation algorithms in online
dating platforms is essential to understand the technological mediation of
romantic connections. Lastly, examining how the theory of "perceived
availability" interacts with other psychological models of attraction
could lead to a more integrated understanding of human mating behaviors. These
diverse research avenues would significantly contribute to our knowledge of
contemporary mate selection processes and their societal implications.
6. References
Association for the
Advancement of Artificial Intelligence (Ed.). (2018). Proceedings of the
Twelfth International AAAI Conference on Web and Social Media: ICWSM: 25-28
June 2018, Stanford, California, USA. International AAAI Conference on Web
and Social Media, Palo Alto, California. AAAI Press.
Brannan,
D., & Mohr, C. D. (2018). Love, friendship, and social support. Noba
textbook series: Psychology. Champaign, IL: DEF publishers.
Buss,
D. M., & Schmitt, D. P. (1993). Sexual Strategies Theory: An evolutionary
perspective on human mating. Psychological Review, 100(2),
204-232. https://doi.org/10.1037/0033-295X.100.2.204
Fernández,
A., Garcia, S., Herrera, F., & Chawla, N. V. (2018). SMOTE for learning
from imbalanced data: Progress and challenges, marking the 15-year anniversary.
Journal of artificial intelligence research, 61, 863-905.
Fisman,
R., Iyengar, S. S., Kamenica, E., & Simonson, I. (2006). Gender Differences
in Mate Selection: Evidence From a Speed Dating Experiment. The Quarterly
Journal of Economics, 121(2), 673-697.
https://doi.org/10.1162/qjec.2006.121.2.673
Hayashi,
T., Mawalim, C. O., Ishii, R., Morikawa, A., Fukayama, A., Nakamura, T., &
Okada, S. (2023). A Ranking Model for Evaluation of Conversation Partners Based
on Rapport Levels. IEEE Access, 11, 73024-73035.
https://doi.org/10.1109/ACCESS.2023.3287984
Joshi,
A., Saggar, P., Jain, R., Sharma, M., Gupta, D., & Khanna, A. (2021).
CatBoost—An Ensemble Machine Learning Model for Prediction and Classification
of Student Academic Performance. Advances in Data Science and Adaptive
Analysis, 13(03n04), Article 03n04.
https://doi.org/10.1142/S2424922X21410023
Kleinerman,
A., Rosenfeld, A., Ricci, F., & Kraus, S. (2018). Optimally balancing
receiver and recommended users’ importance in reciprocal recommender systems. Proceedings
of the 12th ACM Conference on Recommender Systems, 131-139.
https://doi.org/10.1145/3240323.3240349
Lundberg,
S. (2018). SHAP. API Reference. https://tinyurl.com/yhcdt2w8
Lundberg,
S., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model
Predictions. https://doi.org/10.48550/ARXIV.1705.07874
McFarland,
D. A., Broska, D., Prabhakaran, V., & Jurafsky, D. (2024). Coming into
relations: How communication reveals and persuades relational decisions. Social
Networks, 79, 57-75. https://doi.org/10.1016/j.socnet.2024.05.003
Mukhopadhyay,
S. (2018). Advanced Data Analytics Using Python. Apress.
https://doi.org/10.1007/978-1-4842-3450-1
Pincay
Ponce, J. I. (2023). Análisis
de datos educativos aplicado en el estudio de la incidencia de factores
socioeconómicos en el rendimiento escolar [Doctor en Ciencias Informáticas, Universidad Nacional de
La Plata]. https://doi.org/10.35537/10915/156471
Pincay Ponce, J. I., De
Giusti, A. E., Sánchez Andrade, D. A., & Figueroa Suárez, J. A. (2024).
CatBoost: Aprendizaje automático de conjunto para la analítica de los factores
socioeconómicos que inciden en el rendimiento escolar. Revista
Iberoamericana de Tecnología en Educación y Educación en Tecnología, 38,
e3. https://doi.org/10.24215/18509959.38.e3
Pincay-Ponce, J.,
Sánchez-Andrade, D., Caicedo-Ávila, I., & Macías-Valencia, D. (2020,
noviembre 27). Clasificación de pacientes según su posibilidad de adquirir
Diabetes Mellitus empleando algoritmos de Machine Learning. IV Congreso
Internacional Tecnologías de la Información y Computación (CITIC 2020),
Calceta, Ecuador. https://tinyurl.com/yve333v7
Prokhorenkova,
L., Gusev, G., Vorobev, A., Dorogush, A. V., & Gulin, A. (2019). CatBoost:
Unbiased boosting with categorical features (arXiv:1706.09516; Número
arXiv:1706.09516). arXiv. http://arxiv.org/abs/1706.09516
Regan,
P. C. (1998). Minimum Mate Selection Standards as a Function of Perceived Mate
Value, Relationship Context, and Gender. Journal of Psychology & Human
Sexuality, 10(1), 53-73. https://doi.org/10.1300/J056v10n01_04
Sharabi,
L. L., & Dorrance-Hall, E. (2024). The online dating effect: Where a couple
meets predicts the quality of their marriage. Computers in Human Behavior,
150, 107973. https://doi.org/10.1016/j.chb.2023.107973
Van
den Broeck, G., Lykov, A., Schleich, M., & Suciu, D. (2022). On the
Tractability of SHAP Explanations. Journal of Artificial Intelligence
Research, 74, 851-886. https://doi.org/10.1613/jair.1.13283
Weigard,
A., & Spencer, R. J. (2023). Benefits and challenges of using logistic
regression to assess neuropsychological performance validity: Evidence from a
simulation study. The Clinical Neuropsychologist, 37(1), 34-59.
https://doi.org/10.1080/13854046.2021.2023650
Ye,
Y., Ni, K., Jing, F., Zhou, Y., Tang, W., & Zhang, Q. (2024).
Model-Informed Targeted Network Interventions on Social Networks Among Men Who
Have Sex With Men in Zhuhai, China. IEEE Transactions on Computational
Social Systems, 11(1), 238-246.
https://doi.org/10.1109/TCSS.2022.3216756
Zheng,
X., Zhao, G., Zhu, L., Zhu, J., & Qian, X. (2022). What You Like, What I
Am: Online Dating Recommendation via Matching Individual Preferences with
Features. IEEE Transactions on Knowledge and Data Engineering, 1-1.
https://doi.org/10.1109/TKDE.2022.3148485