Analyzing Bike Purchase Predictions Using Machine Learning
Yuzhe Zhang
a
School of Statistics and Mathematics, Zhejiang Gongshang University, Hangzhou, China
Keywords: Bike, Purchase, Machine Learning.
Abstract: As environmental awareness and the promotion of healthy lifestyles continue to rise globally, understanding
the determinants of consumer decisions regarding bicycle purchases has become increasingly important.
Using four machine learning models - Random Forest (RF), Decision Tree (DT), eXtreme Gradient Boosting
(XGBoost), and Logistic Regression (LR) - this study examines the variables influencing bicycle buying
behavior. The experimental results demonstrate that age, income, and car ownership consistently emerged as
significant predictors of bicycle purchasing behavior across all models. Among these, RF and XGBoost
exhibited the best performance in predicting bicycle purchases, with higher accuracy and robustness. The
findings contribute to both theoretical advancements and practical applications, offering valuable insights for
businesses and policymakers aiming to promote cycling as a sustainable mode of transportation. Furthermore,
this study provides a comprehensive framework for understanding the key factors driving consumer decisions,
suggesting that future research should explore hybrid models and additional socioeconomic and cultural
variables for more accurate predictions.
1 INTRODUCTION
The rapid increase in environmental awareness and
the promotion of healthy lifestyles have led to a
significant rise in bicycle purchases worldwide
(Gössling, 2020). This trend is not only evident in
developed countries, where cycling infrastructure and
urban planning are often more supportive of cyclists
(Fishman & Cherry, 2016), but also in emerging
markets where bicycles are becoming an increasingly
popular mode of transport due to their affordability
and environmental benefits (Zhao et al., 2022).
Understanding the factors that influence consumers’
decisions to purchase bicycles is essential for both
theoretical advancements and practical applications
(ZunigaGarcia et al., 2018).
In order to examine bicycle buying behavior, this
study uses four machine learning models: Random
Forest (RF), Decision Tree (DT), eXtreme Gradient
Boosting (XGBoost), and Logistic Regression (LR)
(Sun et al., 2019). These models were chosen for their
varied strengths in handling different types of data
and providing diverse insights. RF and XGBoost are
ensemble methods known for their robustness and
accuracy, especially in managing complex datasets
a
https://orcid.org/0009-0008-2138-3326
with nonlinear relationships. DT models offer
interpretability, allowing clear visualization of
decision rules, while LR provides a straightforward
approach to understanding linear relationships
between variables and the target outcome.
This study aims to explore the key factors
influencing bicycle purchase decisions by comparing
the performance of these four models. The research
objective is to identify the most significant predictors
of bicycle purchases, providing insights that can
inform both market strategies and policy
recommendations aimed at promoting cycling as a
sustainable mode of transportation.
2 LITERATURE REVIEW
2.1 Research on Bicycle Purchase
Behavior Prediction
In recent years, research on predicting bicycle
purchase behavior has gained considerable attention,
especially in the context of rising environmental
awareness and the promotion of healthy lifestyles.
Studies show that socioeconomic characteristics,
274
Zhang, Y.
Analyzing Bike Purchase Predictions Using Machine Learning.
DOI: 10.5220/0013214700004568
In Proceedings of the 1st International Conference on E-commerce and Artificial Intelligence (ECAI 2024), pages 274-279
ISBN: 978-989-758-726-9
Copyright © 2025 by Paper published under CC license (CC BY-NC-ND 4.0)

personal preferences, and environmental factors
significantly influence consumer decision-making.
For example, Gössling (2020) emphasized the
importance of redistributing road space in cities,
highlighting the role of policy interventions in
reducing car dependency and promoting cycling
through better-designed bike lanes. This aligns with
Fishman and Cherry (2016), who found that the
introduction of e-bikes could significantly alter urban
transport patterns.
Similarly, Zhao et al. (2022) explored the
relationship between urban expansion and cycling
behavior in Chinese cities, noting that as urbanization
progresses, cycling becomes more prevalent but is
still significantly shaped by urban planning. Zuniga-
Garcia et al. (2018) compared high- and low-cycling
countries, stressing the importance of infrastructure
and sociocultural factors in promoting cycling. These
studies suggest that improving infrastructure,
adjusting transportation policies, and addressing
cultural factors are key to increasing bicycle
purchases.
In Beijing, Sun et al. (2019) found that commuting
distance, health awareness, and socioeconomic
background significantly impact bicycle purchasing
behavior. Their research indicated that higher-income
groups are more likely to purchase bicycles,
particularly as a means of healthy and
environmentally friendly transportation.
Comprehending these variables is essential for
formulating focused market approaches and policy
measures.
2.2 Application of Machine Learning
Models in Economics
Machine learning techniques have become widely
used in economic data analysis due to their ability to
handle complex datasets. Feng et al. (2021) applied
the RF model to analyze pedestrian injury severity
under different weather conditions, demonstrating the
model's strong capability in handling multi-
dimensional data and capturing non-linear
relationships. The study showed that RF excels in
capturing complex interactions between features,
making it ideal for tasks such as credit risk evaluation
and market prediction.
In terms of DT models, Wibowo and Wibowo
(2020) investigated their performance in classifying
credit scores. DT models, known for their simplicity
and interpretability, are frequently used in economic
decision support systems, especially in scenarios
requiring transparent decision paths, such as customer
segmentation and fraud detection.
Due to its effectiveness in managing complicated
feature interactions and massive amounts of data,
XGBoost has become more and more well-known in
recent years. In a comparative study of XGBoost,
Agrawal and Maurya (2020) discovered that the
model is very good at handling non-linear features
and avoids overfitting, which makes it ideal for
applications like credit scoring and financial market
prediction.
3 MODEL DESCRIPTION
3.1 Random Forest
Random Forest (RF) is an ensemble learning
technique (Feng et al., 2021) that generates several
decision trees during the training phase. For
classification tasks, the output is the most frequent
class prediction, while for regression tasks, the output
is the average prediction.
According to Feng et al. (2021), RF is an
ensemble learning technique that builds several DTs
during training and produces the mean prediction
(regression) or mode of the classes (classification) of
the individual trees. The classification task is shown
in Equation (1).
𝑦 = 𝑚𝑜𝑑𝑒

𝑇


𝑥





1

The regression task is shown in Equation (2).
𝑦 =
1
𝐵
𝑇


𝑥




2

Here T

(x) rdenotes the prediction from the b-th
tree for input x, while B is the total number of trees.
b is the index ranging from 1 to B, representing each
individual tree in the RF.
The RF algorithm, introduced by Breiman, is
designed to improve the accuracy and robustness of
DTs by aggregating the results of multiple trees,
thereby reducing overfitting and enhancing
generalization. RFs are particularly effective for large
datasets with high dimensionality and complex
feature interactions, making them a popular choice in
economics for tasks such as credit risk evaluation,
stock market prediction, and consumer behavior
analysis. The capacity of RFs to handle both
numerical and categorical variables and to provide
feature importance estimates, which aid in the
Analyzing Bike Purchase Predictions Using Machine Learning
275

interpretation of model results, is one of their main
advantages.
3.2 Decision Tree
According to the value of the input features, a
Decision Tree (DT) model divides the data into
subsets (Wibowo. A & Wibowo. P. B, 2020). The
Gini impurity, which calculates the likelihood that an
element selected at random would receive an
inaccurate label if its label were assigned at random
based on the dataset's label distribution is defined as
Equation (3).
𝐺𝑖𝑛𝑖 =1−𝑝




(
3
)
p

represents the proportion of observations
belonging to class i, and n is the number of classes.
DTs are one of the simplest yet powerful
predictive modeling approaches. They are highly
interpretable because they represent decisions and
their possible consequences in a tree-like structure.
This method is particularly useful when the primary
goal is to gain insights into the decision-making
process rather than just prediction. Due to its capacity
to manage intricate relationships between variables
and produce precise, actionable rules, DT is
frequently employed in economic applications for
purposes such as consumer segmentation, fraud
detection, and identifying critical drivers of economic
outcomes.
3.3 Extreme Gradient Boosting
A gradient-boosting system called Extreme Gradient
Boosting (XGBoost) creates DTs in a sequential
fashion with the goal of fixing mistakes committed by
earlier trees. Equation (4) provides a definition for the
objective function.
𝐿
(
𝜙
)
= 𝑙


(
𝑦

, 𝑦

)
+ 𝛺


(
𝑓

)(
4
)
Define l
(
y

,y

)
as the loss for prediction y

, and
Ω
(
f

)
as the regularization term for tree k. It was
developed by Chen and Guestrin (2020), XGBoost is
known for its efficiency, speed, and scalability
(Lundberg & Lee, 2020). It has become one of the
most popular machine learning algorithms in recent
years, particularly in data science competitions and
realworld applications that require high predictive
accuracy (Agrawal & Maurya, 2020). XGBoost’s
ability to handle missing data, capture complex
patterns, and prevent overfitting through
regularization makes it ideal for economic modeling
tasks, such as predicting financial market trends,
credit scoring, and analyzing large-scale economic
data (Chen & Guestrin, 2020).
3.4 Logistic Regression
Logistic Regression (LR) is a linear model for binary
classification. The logistic function, which models
the probability that a given input point belongs to a
particular class, is given by Equation (5).
𝑃
(
𝑌 =1|𝑋
)
=
1
1+𝑒

(






⋯



)
(
5
)
𝛽

is the intercept, 𝛽

… 𝛽

are the coefficients,
and x

are the predictor variables.
Because of its ease of use and efficiency in binary
classification problems, LR is one of the most popular
statistical models. It is easily interpreted and simple
to apply because it assumes a linear relationship
between the input variables and the response
variable's log odds. In economic research, LR is often
employed to model binary outcomes such as purchase
decisions, default risk, and employment status. Its
ability to provide insights into the influence of
predictor variables and its relatively low
computational cost makes it a preferred choice for
many applied econometric analyses.
4 EMPIRICAL ANALYSIS
4.1 Model Results Comparison
As shown in Table 1, in this study, four machine
learning models - RF, DT, XGBoost, and LR - were
employed to analyze bicycle purchase behavior. The
models predict whether a user will purchase a bicycle,
with the positive class representing "purchase" and
the negative class representing "no purchase." The
performance of each model is summarized in terms of
accuracy, precision, recall, and F1-score.
Among the models, RF demonstrated the highest
overall performance, with an accuracy of 71%,
followed closely by XGBoost at 69.5%. In contrast,
the DT and LR models performed less effectively,
achieving accuracies of 63% and 56%, respectively.
It is worth noting that RF and XGBoost performed
better in both precision and recall, particularly for the
positive class (bicycle purchase), indicating their
superior ability to identify actual purchasing
ECAI 2024 - International Conference on E-commerce and Artificial Intelligence
276

Table 1: Model Performance Comparison.
Model Accurac
y
Precision
(
0
)
Precision
(
1
)
Recall
(
0
)
Recall
(
1
)
F1
(
0
)
F1
(
1
)
RF 0.7100 0.7182 0.7000 0.7453 0.6702 0.7315 0.6848
DT 0.6300 0.6455 0.6111 0.6698 0.5851 0.6574 0.5978
XGBoost 0.6950 0.7228 0.6667 0.6887 0.7021 0.7053 0.6839
LR 0.5600 0.5833 0.5326 0.5943 0.5213 0.5888 0.5269
Table 2: Feature Importance Across Models.
Feature RF DT XGBoost LR
Income 0.1220 0.1404 0.0335 0.4216
Children 0.0959 0.0786 0.0309 -0.2141
Cars 0.0975 0.0976 0.0581 -0.6129
Age 0.2078 0.1982 0.0356 0.1013
Martial Status
_
Sin
g
le 0.0454 0.0538 0.0327 0.3751
Gender
_
Male 0.0507 0.0493 0.0311 0.0844
Education
_
Graduate De
g
ree 0.0168 0.0223 0.0397 -0.1863
Education_High School 0.0220 0.0320 0.0324 0.0791
Education_Partial College 0.0235 0.0126 0.0193 -0.0255
Education_Partial High School 0.0148 0.0088 0.1208 -0.1024
Occu
p
ation
_
Mana
g
ement 0.0117 0.0093 0.0137 -0.0753
Occu
p
ation
_
Manual 0.0125 0.0054 0.0251 -0.0827
Occu
p
ation
_
Professional 0.0198 0.0188 0.0371 0.1073
Occupation_Skilled Manual 0.0195 0.0219 0.0341 -0.0659
Home Owner_Yes 0.0358 0.0355 0.0341 0.1844
Commute Distance
_
1-2 Miles 0.0333 0.0419 0.0431 -0.0991
Commute Distance
_
2-5 Miles 0.0309 0.0409 0.0336 -0.0855
Commute Distance
_
5-10 Miles 0.0311 0.0414 0.0439 -0.3285
Commute Distance_More than 10 Miles 0.0184 0.0084 0.0292 -0.3770
Region_North America 0.0271 0.0119 0.0351 0.0217
Region_Pacific 0.0222 0.0350 0.0938 0.3419
A
g
e Brackets
_
Middle A
g
e 0.0255 0.0301 0.1432 0.2280
A
g
e Brackets
_
Ol
d
0.0157 0.0059 0.0000 0.0098
behavior. Although LR provides strong linear
interpretability, its performance is hindered by its
inability to capture complex relationships between
variables, resulting in lower accuracy and recall rates.
4.2 Feature Importance Analysis
As shown in Table 2, in this study, based on the model
results, this paper conducted a detailed analysis of the
key features influencing bicycle purchases. Below,
this paper focuses on the most important features,
considering both model performance and their real-
world implications.
The positive category represents users who buy
bicycles, and the negative category represents users
who do not buy bicycles. Understanding which
characteristics make users more inclined to buy or not
buy bicycles, can provide relevant policymakers and
enterprises with a reference to help them better
promote bicycle sales or develop related marketing
strategies.
Income emerged as a critical factor in determining
bicycle purchases across all models. In the LR model,
the income coefficient is 0.421572, indicating a
strong positive correlation between higher income
levels and the likelihood of purchasing a bicycle. This
aligns with real-world consumer behavior, where
higher-income individuals are more likely to view
bicycles as part of a healthy lifestyle or as an
environmentally friendly mode of transport.
Furthermore, higher income levels enable the
purchase of higher-end bicycles, which could
contribute to this trend.
The number of cars owned by a household or
individual showed a negative correlation with bicycle
purchases in multiple models. For instance, in LR, the
coefficient for car ownership is 0.612922, suggesting
that owning more cars significantly decreases the
likelihood of purchasing a bicycle. This is consistent
with the notion that car ownership reduces the need
for bicycles as a primary mode of transport. Policies
aimed at reducing car dependency or promoting
Analyzing Bike Purchase Predictions Using Machine Learning
277

bicycles as a viable transportation alternative may
help increase bicycle sales.
Age was identified as one of the most important
predictors in the RF and DT models. Older
individuals were more likely to purchase bicycles, a
finding that may be explained by health
considerations. As people age, they often become
more concerned with maintaining physical health,
and cycling is increasingly viewed as an accessible
and beneficial form of exercise. From a practical
perspective, marketing campaigns targeting older
demographics that emphasize the health benefits of
cycling could further stimulate bicycle sales in this
segment.
4.3 Key Factors in Bicycle Purchase
Behavior
The results of this study highlight the key factors
influencing bicycle purchase behavior as identified
by the four models. Age, income, and car ownership
consistently emerged as significant predictors
regardless of the model used. These insights provide
valuable information for businesses seeking to target
potential customers, as well as policymakers looking
to promote cycling as a sustainable transportation
option.
For example, RF and XGBoost both identified age
as a crucial factor, indicating that older individuals
are more likely to purchase bicycles. This finding
suggests that marketing strategies could be designed
specifically to appeal to older demographics by
emphasizing the health benefits of cycling and
positioning bicycles as an effective means of
maintaining physical well-being.
4.4 Policy and Business Implications
Similarly, income played a vital role in predicting
bicycle purchases across all models. Higher-income
individuals are more likely to buy bicycles, possibly
due to their greater financial capacity to invest in
high-quality bicycles and a lifestyle that promotes
sustainable transportation. Businesses could focus on
premium pricing strategies or offer advanced bicycle
features targeting affluent consumers.
Car ownership, which showed an inverse
relationship with bicycle purchases, highlights an
area of potential policy intervention. Reducing car
dependency and offering incentives for cycling - such
as building more cycling infrastructure or providing
tax breaks for bicycle purchases - could increase
bicycle sales. This insight helps policymakers create
effective strategies for promoting healthier, more
sustainable urban environments.
5 CONCLUSIONS
The rapid rise in environmental awareness and the
promotion of healthier lifestyles have significantly
contributed to the increase in bicycle purchases
globally. Understanding the factors driving consumer
decisions to purchase bicycles is crucial for both
theoretical research and practical applications, as it
offers valuable insights for businesses and
policymakers alike. This study applied four machine
learning models—RF, DT, XGBoost, and LR—to
analyze bicycle purchase behavior. Each model
demonstrated unique advantages, with RF and
XGBoost showing the most robust predictive
performance. Key features such as age, income, and
car ownership were identified as significant
predictors, reflecting real-world consumer trends and
socioeconomic factors. This study is not without
limits, though. Initially, the models employed were
restricted to a particular group of variables, which
might not adequately represent the intricacy of
customer behavior. Additionally, the data used for
model training may not account for regional or
cultural differences in bicycle purchasing patterns.
These limitations suggest that further research should
incorporate additional variables, such as lifestyle
choices, environmental policies, and urban
infrastructure, to better understand consumer
behavior. Future studies should investigate hybrid
models, which combine the best features of various
machine learning approaches to increase
interpretability and accuracy. Expanding the scope of
analysis to include broader socioeconomic and
cultural factors will allow for a more comprehensive
understanding of bicycle purchase behavior.
Additionally, investigating the role of infrastructure
development, such as bike lanes and public cycling
initiatives, could provide further insights into how
urban planning can promote cycling as a primary
mode of transportation.
REFERENCES
Agrawal, S., Maurya, A. K., 2020. A comparative analysis
of XGBoost. International Journal of Scientific &
Technology Research, 9(2), 132-136.
Chen, T., Guestrin, C., 2020. XGBoost: A Scalable Tree
Boosting System. ACM Transactions on Knowledge
Discovery from Data, 14(1), 1-24.
ECAI 2024 - International Conference on E-commerce and Artificial Intelligence
278

Feng, J., Wu, Y., Xu, Z., 2021. RF-based pedestrian injury
severity analysis for different weather conditions.
Accident Analysis & Prevention, 149, 105870.
Fishman, E., Cherry, C., 2016. E-bikes in the mainstream:
Reviewing a decade of research. Transport Reviews,
36(1), 72-91.
Gössling, S., 2020. Why cities need to take road space from
cars—and how this could be done. Journal of Urban
Design, 25(4), 443-448.
Lundberg, S. M., Lee, S. I., 2020. A Unified Approach to
Interpreting Model Predictions. Advances in Neural
Information Processing Systems, 33.
Sun, G., Li, W., Zeng, C., 2019. Bicycle commuting
behavior in Beijing: Lessons from the commuters.
Transport Policy, 82, 58-67.
Wibowo, A., Wibowo, P. B., 2020. Performance of
Decision Tree algorithms in classifying credit scoring.
Journal of Physics: Conference Series, 1566(1),
012033.
Zhao, P., Lu, B., Ge, J., 2022. Urban expansion and cycling
behavior: Evidence from Chinese cities. Journal of
Transport Geography, 100, 103284.
ZunigaGarcia, N., Nazelle, A., Nieuwenhuijsen, M. J.,
2018. Cycling in cities: Key lessons learned from high-
and low-cycling countries. International Journal of
Environmental Research and Public Health, 15(4),
556.
Analyzing Bike Purchase Predictions Using Machine Learning
279