The Prediction of Feedback with Demographics & Locations of Users

in Online Food Delivering Based on Machine Learning Models

Junyi Hu

Institute of Science of Mathematics, Nanjing Normal University, Nanjing, China

Keywords: Prediction, Demographics & Locations, Online Food Delivering, Machine Learning Models.

Abstract: The online food delivery (OFD) becomes a hit in the previous decade. However, the OFD companies face

different challenges in reality. With the data providing demographic features and locations, this research

makes use of data learning models to predict the feedback of the OFD users, which implies the probability of

becoming a long-term user. This research takes ANN neural network, Decision Tree, Random Forest into

consideration, and make comparison of the performance of them, and all the models reach the accuracy above

80%. The geographic feature of the users is considered in an abstract map, and the geographic center of the

user with negative feedback is to the north-west of the geographic center of all the users. In addition, feature

analysis reveals the age of user is the feature with most importance, followed by family size. The performance

of the models is estimated by different evaluating indicators, such as confusion matrix, f-measure and

accuracy.

1 INTRODUCTION

With the advancement of information technology, the

widespread application of OFD services is evident.

However, not everyone readily embraces this new

trend. Therefore, investigating the primary user

demographics of OFD services is a worthwhile

research endeavor. Factors such as Information

Quality (IQ) and Promotions (PRO) significantly

influence whether an individual becomes an online

food delivery user and their satisfaction with the

service. Accurately predicting the likelihood of

someone becoming a potential user of online food

delivery services based on individual factors is crucial

(Moroz and Polkowski, 2016; Prasetyo and Tanto,

2021 & Tan and Kim, 2021). Additionally,

identifying which features are the most influential

factors is of utmost importance (Shukla and

Deshpande, 2023 & Wang et al., 2018). It can help

the retailers and OFD platforms figure out the most

likely long-term users and make better programming

and advertising.

Previous research on OFD services has primarily

focused on macro-level aspects, such as consumer

attitudes towards these services and the structural

components of the OFD system, including factors like

convenience motivation and the perceived usefulness

after usage (Yeo et al., 2017). However, these studies

often overlook the nuanced details, particularly those

concerning individual users within the OFD service

ecosystem. They tend to treat OFD users as a

monolithic entity, disregarding the intricate internal

structure of consumer communities, which is a

critical aspect addressed in this current research.

The neglect of individual user experiences and

behaviors within the broader OFD consumer base can

lead to a lack of understanding of the diverse

motivations, preferences, and challenges faced by

different user segments. For instance, some users

might prioritize speed of delivery, while others might

be more concerned with the quality of food or the

sustainability of the delivery process. Similarly, the

frequency of OFD usage, the types of meals ordered,

and the time of day when orders are placed can vary

significantly among users, reflecting diverse

lifestyles and needs.

Moreover, the socio-demographic characteristics

of users, such as age, income level, and geographical

location, can influence their interactions with OFD

services. Younger, more tech-savvy consumers might

be more inclined to use mobile apps for ordering,

while older users might prefer traditional methods

like phone calls. Urban dwellers might have access to

a wider variety of restaurants and faster delivery

times compared to those in rural areas.

Hu, J.

The Prediction of Feedback with Demographics & Locations of Users in Online Food Delivering Based on Machine Learning Models.

DOI: 10.5220/0012982500004601

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

In Proceedings of the 1st International Conference on Innovations in Applied Mathematics, Physics and Astronomy (IAMPA 2024), pages 13-21

ISBN: 978-989-758-722-1

This paper investigates the features that make an

individual a potential consumer of OFD services. By

employing supervised machine learning models and

incorporating feedback from OFD users, we aim to

delineate the profile of potential OFD users.

Recognizing the significant impact of dataset quality

on accuracy based on previous research, this study

utilizes a dataset that includes geographical locations

and demographic information related to online food

delivery. Extensive efforts were made to portray a

general image of OFD service users using this data.

Multiple models were implemented in this study, and

a comprehensive comparison of these models was

conducted to leverage their strengths and address

their weaknesses. Additionally, we placed particular

emphasis on geographical factors such as longitude

and latitude, as well as demographic factors like

household size, education level, and monthly income.

To present the findings clearly and intuitively,

machine learning models and data visualization

techniques were emphasized.

The dataset comprises rich independent features

such as latitude and longitude, gender, occupation,

household size, and individual feedback. The aim of

this research is to analyze relevant data through

machine learning models to comprehensively depict

the profile of potential OFD users, thereby

uncovering common characteristics and needs within

these user groups. This approach not only provides

guidance for OFD platforms to customize their

services effectively and improve customer

satisfaction but also offers insights for refining

marketing strategies and user interface designs. By

thoroughly analyzing these factors, we hope to

enhance the understanding of diverse user group

needs and thus support the development of the online

food delivery industry.

The remainder part of this paper is constructed

below. The second section is about the related work

about the OFD service. Section 3 is mainly about the

research method and detail of the dataset. The

machine learning model applied is discussed in

Section 4. Section 5 is about the result and the

discussion of this experiment. Finally, we discuss the

conclusion of the result and future work about OFD

services.

2 RELATED WORKS

The satisfaction of the user of online food delivery

system is determined by different factors. Previous

studies on factors affecting satisfaction with online

food delivery services have spanned a wide range of

research areas. Machine learning witnessed a vast

progress in the last decades, applied in extensive

aspects of researches. Common machine learning

methods include random forest, XGBoost, decision

trees, artificial neural networks, Bayesian networks,

and so on. (Yeo et al., 2017) Some studies trained

their models on datasets and employed various

machine learning models for prediction. Their

research achieved high accuracy. Huycock Tan

(2021) initially implemented a linear regression

model, providing insights into online food delivery

users. Their model focused on online food delivery

services during the COVID-19 pandemic, illustrating

factors that positively impact OFD user satisfaction.

Another study by Janmejay (2024) applied a decision

tree classification model, uncovering preferences of

OFD users, offering better decision guidance for

online food delivery aggregators. SVM and Ridge

regression was applied in the research of Wang, W.M.

(2018).

So far, the world of auto machine learning models

and deep learning models has enlarged greatly. For

example, PLS-ANN (Foo et al., 2018) and Bi-LSTM

(Tam and Tanriöver, 2021). They have been

successfully applicated into a wide range of fields and

gain outstanding results.

3 METHOD

In this research, dataset is preprocessed into the form

which the machine learning models and neural

network are easy to manipulate on. machine learning

techniques are implemented to predict the likelihood

of potential user based on the features given.

Optimization are conduct to improve the performance

of the models, especially for the neural network.

Based on the algorithm of random forest, the

influence of different features on the result are

conducted (Figure 1).

Figure 1: Research workflow (Picture credit: Original).

IAMPA 2024 - International Conference on Innovations in Applied Mathematics, Physics and Astronomy

Table 1: Data Info.

Gender

Marital

Status

ccupatio

Monthly Income

Educational

Qualifications

Family

size

latitude

ongitud

in cod

utpu

eedbac

0 20 Female Single Student No Income Post Graduate 4 12.9766 77.5993 560001 Yes Positive

1 24 Female Single Student Below Rs.10000 Graduate 3 12.977 77.5773 560009 Yes Positive

2 22 Male Single Student Below Rs.10000 Post Graduate 3 12.9551 77.6593 560017 Yes Negative

... ... ... ... ... ... ... ... ... ... ... ... ...

387 23 Male Single Student No Income Post Graduate 5 12.8988 77.5764 560078 Yes Positive

3.1 Data Analysis

Before the data preprocessing, having a rough

understanding about the dataset by data analysis is

crucial. The features and the data type are shown in

table 1.

3.2 Data Preprocessing

Before building and training the data in the model, data

preprocessing is necessary. To transform the data into

the form which is easily implemented on, the labels of

features, such as occupation, marital status are turned

into categories labelled with number.

Focusing on the importance of different features,

we need to transform the scale of different features

with the standard score. It dismisses the value of each

figure. Instead, the process of the z-scoring focuses on

the distance of every point and the average of the

feature. It follows the formulas given.





(𝜇=

∑











,𝜎=



∑



𝑥



−𝑢









) (1)

3.3 Location Analysis

The feedback distribution by location is shown in the

figure 2.

The location of the users is the central part of a

small town in India. Firstly, the data show that the

users of OFD are concentrate around the central point

with Latitude 12.97

。

N and Longitute77.60

。

Density of the users decrease with distance away

from the central point. In the plot, we can figure out

the majority of the users with negative feedback lie in

the area whose longitude in the range of 77.55

。

E and

77.62

。

E Latitude between 12.92

。

N and 13.07

。

N, and

the central point of users with negative feedback is

(12.99

。

N,77.57

。

E). In addition, the central point of

users with positive feedback is (12.96

。

N ,77.64

。

E),

to the south-east of the central point of negative-

feedback-uses.

4 MODEL DESIGN

4.1 Neural Network

A neural network is a model imitating the structure

and effect of the human brain. It consists of

interconnected vertices, called neurons(nodes),

organized into layers (Agatonovic-Kustrin and

Beresford, 2000). Each layer is composed by multiple

artificial neurons or processing elements, connected

with weights(coefficients). A neural network

typically consists of input layer, hidden layer and

output layer.

It is easy to see the input layer takes in the data,

and the output layer gives the result. The activate

function, typically non-linear is applied to deal with

the data given by the neurons in the previous layer,

and transform the data into different signals, which is

not seen by the outside. Such process is implemented

in the hidden layers. The model training happens in

the form of adjusting the weight of the connection

between different neurons and layers (Figure 3).

In this research, the ANN model takes for layers

of neurons, with the input layer with 3 neurons and

the output layer with 2 neurons. The hidden layer

contains 16 neurons each. Every layers are connected

by ReLU activation function.

The Prediction of Feedback with Demographics & Locations of Users in Online Food Delivering Based on Machine Learning Models

Figure 2: Location (Picture credit: Original).

Figure 3: Neural Network (Agatonovic-Kustrin and Beresford, 2000).

In this research, there are three fully connected

layers in the ANN neural network. The ReLU

activation function is applied between different layers

and applying drop out after each activation function.

BCEwithLogistsLOSS play the role of loss function,

which combines a sigmoid activation function and the

binary cross entropy loss. The optimizer is ADAM

with learning rate 0.0001.

In ANN the optimization is about the activation

function. The sigmoid function, arctan function and

ReLU function were used. Because of the similarity

of the sigmoid function and arctan activation

function, the results are very similar. The ReLU

activation function perform best.

4.2 Decision Tree

Decision tree is a basic model in machine learning,

based on classification and regression. In the decision

tree there is an important value that describes the

degree of internal difference and uncertainty of a

category, which is typically called entropy. The

smaller the entropy, the better the effect of

classification. The value of entropy is given by the

following formula

H(x)=− 𝑝



∗log 𝑝









(2)

The decision tree model divides the dataset into

several parts by one of the features. Then the model

computes the entropy and the information gain ratio

of different categories. The feature with the highest

information gain ratio is applied first. With the first

feature, the model can make the first decision, which

classify the dataset into several categories with the

lowest entropy. Now the first layer of the tree is

finished, but the classification is not precise enough

to make the prediction. In order to improve the

accuracy, the process above is repeated. The model

finds the best feature apart from the feature applied

above. And the second layer is formed. However, if

all the feature were used, the model may have the

problem of overfitting. Consequently, some

optimization such as pruning and make limitation on

the number of layers of the tree.

In this research, the criterion is chosen as the

entropy. The max depth is decided to be 10.

IAMPA 2024 - International Conference on Innovations in Applied Mathematics, Physics and Astronomy

Figure 4: Random Forest (Wang et al., 2018).

At first the decision tree model performed the

worst, for both the feature of the dataset and the

model overfitting. In order to tackle with the problem,

we first applied different algorithms including CART,

C4.5 and ID3. Then we limited the depth in order to

prevent overfitting.

4.3 Random Forest

Random forest is a state-of-art ensemble machine

learning model. The essence of random forest is

generating multiple decision trees (Figure 4). The

model picks up a subset of k-features of the features

give, typically k=log(n), and n is the total number of

the features. The importance of each feature can be

computed in the process of random forest predicting,

which is a great advantage of the model. There are

many classification trees in a random forest. We need

to classify an input sample by inputting it into each

tree for classification. Finally, the mean or majority

vote in classification is the result.

In this research, the random forest is provided

with a set of parameters and is allowed to choose the

parameter to gain the highest accuracy. The result of

the parameter decision is the n-estimator is 1000 and

the max depth is 10.

5 RESULT AND DISCUSSION

This research put eyes on the performance of different

machine learning models on the dataset to make

prediction on the feedback of the user according to

the demographic features and locations. This research

take accuracy, confusion matrix, recall F-measure as

the evaluation criterion. Accuracy is the most

important feature about a prediction model, while

recall ratio and confusion matrix helps the

optimization of the models.

In this research, all models were implemented on

python 3.10.9 environment, with Numpy, Pandas,

Scikit-Learn and Pytorch packages. The hardware

configurations comprise a 3.20 GHz AMD7 5800H

CPU, a RTX 3060 GPU.

5.1 Dataset Analysis

The paper utilizes the online food dataset from

Kaggle. Each input data consists 12 attributes. We

explore the correlation of between the attributes

completely, in order to find out the feature that are

conducive to the predicting of feedback. A heat map

was given in figure 5.

In this figure 6, we can conclude that most users

of online food service gave positive feedbacks. The

age distribution is the young people with age lower

than 25 is in the most favor of the service. Especially

the 29-year-old and 30-year-old population give more

negative feedback than the positive, which is rare in

this research. Male users take a higher proportion than

the female, and they contribute a higher satisfaction

rate. The majority of the marital status of the users is

single, and the students show most interest in the

online food service, particularly post graduate

students.

The Prediction of Feedback with Demographics & Locations of Users in Online Food Delivering Based on Machine Learning Models

Figure 5: Heatmap (Picture credit: Original).

5.2 Model Prediction and Optimization

5.2.1 Confusion Matrix

The confusion matrix of machine learning predicting

model is a crucial criterion of the performance of the

0-1 prediction models. It conveys the information

about the actual data and the prediction intuitively.

The confusion matrices of the models are shown

in figure 7.

5.2.2 Evaluation of the Models

The evaluation of the criterion on the models above

are revealed in the table 2.

In the table we can get the conclusion that the

accuracy of the three models are all above 80%, and

the decision tree get the highest, while ANN is the

lowest. The possible reason of the unsatisfying

performance of ANN is the scale of dataset is too

small. However, in the field of recall, ANN take the

lead, with 90.47% came to the first. Decision tree also

have the highest precision rate. The number of

features of the data is in the performance advantage

zone of Decision Tree model, which is a reason of the

performance of Decision Tree. The differences of F-

measure are at a small extent.

IAMPA 2024 - International Conference on Innovations in Applied Mathematics, Physics and Astronomy

Figure 6: Counting of data (Picture credit: Original).

Figure 7: Confusion matrices (Picture credit: Original).

The Prediction of Feedback with Demographics & Locations of Users in Online Food Delivering Based on Machine Learning Models

Figure 8: feature importance (Picture credit: Original).

Table 2: Models Evaluation.

Model Accuracy Recall Precision

ANN 80.76% 90.47% 87.02%

Random Forest 83.97% 88.54% 88.54%

Decision Tree 87.82% 87.59% 92.30%

5.3 Importance Analysis

The random forest model has a great advantage. The

model brings up multiple decision trees and take the

average of the results of different trees. In this

process, the distribution of each feature on the final

result is shown clearly.

We divide the features into two classes. For the

features with plenty of value, such as age, family size,

are considered in category. However, for some

features such as occupation are considered in the form

of concrete feature. The result is shown in the figure

In the see the age is the feature with highest

importance, followed by family size. The importance

of other features are far lower

6 CONCLUSION AND FUTURE

WORKS

In order to address the issue faced by OFD companies

when determining potential users, this research

utilizes a Kaggle dataset and machine learning

models to predict user feedback based on

demographic characteristics and geographic location.

The study identifies age as the decisive factor in user

feedback, providing guidance to companies on

targeting specific audiences. All models achieved an

accuracy of over 80%, with high overall evaluation

metrics, indicating outstanding research results.

Among the models used, Decision Tree outperformed

others in various indicators, while the ANN

performed moderately. This suggests that Decision

Tree excels in binary prediction tasks, while ANN's

performance is limited due to its two-node output

layer. However, the research has certain limitations,

including a relatively small dataset and

homogenization bias towards users with positive

feedback. Future research should consider larger and

more diverse datasets, as well as explore optimization

strategies for models, particularly for ANN models.

This research also has some limitations. The data

is only about the OFD users in a town in India, which

may lead to the deprivation of universality of the

research. Future studies can widen the scope of

dataset and trying to find the uniform regularity of

OFD users. Additionally, the respective number of

models applied is limited. Future researches can do

some more optimizations and make use of a wider

range of machine learning models to gain a more

precise prediction.

REFERENCES

M. Moroz and Z. Polkowski, "The Last Mile Issue and

Urban Logistics: Choosing Parcel Machines in the

Context of the Ecological Attitudes of the Y Generation

Consumers Purchasing Online", Transp. Res. Procedia,

vol. 16, pp. 378-393, 2016.

Prasetyo, Y.T.; Tanto, H; Factors Affecting Customer

Satisfaction and Loyalty in Online Food Delivery

Service during the COVID-19 Pandemic: Its Relation

IAMPA 2024 - International Conference on Innovations in Applied Mathematics, Physics and Astronomy

with Open Innovation. J. Open Innov. Technol.

Mark. Complex. 2021, 7, 76.

Tan, H., & Eng Kim, V. W. (2021). Examining the Factors

that Influence Consumer Satisfaction with Online

Food Delivery in Klang Valley, Malaysia. The Journal

of Management Theory and Practice

(JMTP), 2(2), 88-95.

Shukla, J., & Deshpande, A. (2023). A decision tree

classifier approach for predicting customer’s

inclination toward use of online food delivery services.

Multidisciplinary Science Journal, 6(5), 2024072

https://doi.org/10.31893/multiscience.2024072

Wang, W.M.; Wang, J.W.; Barenji, A.V.; Li, Zhi; Tsui, Eric

(2018). Modeling of individual customer delivery

satisfaction: an AutoML and multi-agent system

approach. Industrial Management & Data

Systems, IMDS-07-2018-0279

Vincent Cheow Sern Yeo, See-Kwong Goh, Sajad Rezaei,

Consumer experiences, attitude and behavioral

intention toward online food delivery (OFD) services,

Journal of Retailing and Consumer Services,

Volume 35, 2017, Pages 150-162, ISSN 0969-6989,

Pik-Yin Foo, Voon-Hsien Lee, Garry Wei-Han Tan, Keng-

Boon Ooi, A gateway to realising sustainability

performance via green supply chain management

practices: A PLS–ANN approach, Expert Systems

with Applications, Volume 107,2018, Pages 1-14,

S. Tam, R. B. Said and Ö. Ö. Tanriöver, "A ConvBiLSTM

Deep Learning Model-Based Approach for Twitter

Sentiment Classification," in IEEE Access, vol. 9, pp.

41283-41293, 2021, doi: 10.1109/ACCESS.2021.3064

830.

Agatonovic-Kustrin S, Beresford R. Basic concepts of

artificial neural network (ANN) modeling and its

application in pharmaceutical research. Journal of

pharmaceutical and biomedical analysis, 2000, 22(5):

717-727.

The Prediction of Feedback with Demographics & Locations of Users in Online Food Delivering Based on Machine Learning Models