Bachelor of Science in Computer Science

May 2020

Predicting sales using Machine

Learning Techniques

Sai Nikhil Boyapati

Ramesh Mummidi

Faculty of Computing, Blekinge Institute of Technology, 371 79 Karlskrona, Sweden

This thesis is submitted to the Faculty of Computing at Blekinge Institute of Technology in
partial fulfilment of the requirements for the degree of Bachelor of Science in Computer Science.
The thesis is equivalent to 20 weeks of full time studies.

The authors declare that they are the sole authors of this thesis and that they have not used
any sources other than those listed in the bibliography and identified as references. They further
declare that they have not submitted this thesis at any other institution to obtain a degree.

Contact Information:
Author(s):
Sai Nikhil Boyapati
E-mail: [email protected]

Ramesh Mummidi
E-mail: [email protected]

University advisor:
Suejb Memeti
Department of computer science (DIDA)

Faculty of Computing Internet : www.bth.se

Blekinge Institute of Technology Phone : +46 455 38 50 00
SE–371 79 Karlskrona, Sweden Fax : +46 455 38 50 57
 Abstract

Background: Sales forecasting is an important field in the food sector, and it has
recently got immense popularity to boost market operations and productivity due to
new technologies. The industry has traditionally focused on a conventional statisti-
cal model but in the recent years, Machine Learning techniques have received more
attention.
Objectives: This thesis will help to identify the critical features that influence sales
and also an experiment is performed to find the best suitable algorithm for sales
forecasting.
Methods: Machine Learning Algorithms such as Simple Linear Regression, Gradi-
ent Boosting Regression, Support Vector Regression, and Random Forest Regression
were considered in this thesis, which they expected to perform well on the issues. An
experiment is carried out to determine the efficiency of the algorithms.
Results: Algorithms such as Simple Linear Regression, Gradient Boosting Re-
gression, Support Vector Regression, and Random Forest Regression are commonly
known for performing better than others, this has been clearly shown that Random
Forest Regression is the most appropriate algorithm compared to the others.
Conclusions: The Random Forest Regression algorithm performed well after do-
ing all the study when compared with other algorithms. Hence the Random Forest
Regression is considered as the best suitable algorithm for forecasting product sales.

Keywords: Correlation, Machine Learning, Performance Metrics, Sales Forecasting

 Acknowledgments

First and foremost, praises and thanks to God, the Almighty, for His showers of bless-
ings throughout our research work to complete the research successfully. We would
like to express our deep and sincere gratitude to our supervisor Suejb Memeti, for
giving us the opportunity to do research and providing invaluable guidance through-
out this research. Under his guidance it was a great privilege and honour to work
and study. We are extremely appreciative of what he has offered us.
Finally, our special thanks to all the people who have helped us directly or indi-
rectly in completing the research work.

ii
 Contents

Abstract i

Acknowledgments ii

1 Introduction 1
1.0.1 Aims and Objectives . . . . . . . . . . . . . . . . . . . . . . . 2
1.0.2 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.3 Machine Learning Algorithms: . . . . . . . . . . . . . . . . . . 5
1.1.4 Selection of Machine Learning Algorithms . . . . . . . . . . . 6
1.1.5 Selection of Performance Metrics . . . . . . . . . . . . . . . . 7

2 Related Work 8

3 Method 10
3.1 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Experimentation Environment . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Data overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4 Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4.1 Data Correlation Method . . . . . . . . . . . . . . . . . . . . 13
3.5 Feature Importance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.6 Data preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.6.1 Encoding Categorical Values . . . . . . . . . . . . . . . . . . . 15
3.6.2 Stratified K-fold Cross-Validation . . . . . . . . . . . . . . . . 17
3.7 Performance Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.7.1 Accuracy score . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.7.2 Max Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.7.3 Mean Absolute Error . . . . . . . . . . . . . . . . . . . . . . . 18

4 Results 19
4.1 Simple Linear Regressor . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 Gradient Boosting Regressor . . . . . . . . . . . . . . . . . . . . . . . 19
4.3 Support Vector Regressor . . . . . . . . . . . . . . . . . . . . . . . . 22
4.4 Random Forest Regressor . . . . . . . . . . . . . . . . . . . . . . . . 23
4.5 Feature Importance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.6 Evaluation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

iii
5 Analysis and Discussion 27
5.1 Comparative analysis of Performance Metrics . . . . . . . . . . . . . 27
5.1.1 Average Accuracy Score . . . . . . . . . . . . . . . . . . . . . 27
5.1.2 Average Mean Absolute Error . . . . . . . . . . . . . . . . . . 28
5.1.3 Average Max Error . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.4 Validity Threats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.4.1 Internal Validity . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.4.2 External Validity . . . . . . . . . . . . . . . . . . . . . . . . . 30

6 Conclusions and Future Work 31

6.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

References 32

A Appendix 36

iv
 List of Figures

1.1 Types of Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . 4

3.1 Dataset description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.2 Dataset summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3 Correlation values . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.4 Heat map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.5 Before One Hot Encoding . . . . . . . . . . . . . . . . . . . . . . . . 16
3.6 After One Hot Encoding . . . . . . . . . . . . . . . . . . . . . . . . . 16

4.1 Accuracy Box plot for Simple Linear Regressor . . . . . . . . . . . . . 20

4.2 MAE Box plot for Simple Linear Regressor . . . . . . . . . . . . . . . 20
4.3 ME Box plot for Simple Linear Regressor . . . . . . . . . . . . . . . . 20
4.4 Accuracy Box plot for Gradient Boosting Regressor . . . . . . . . . . 21
4.5 MAE Box plot Gradient Boosting Regressor . . . . . . . . . . . . . . 21
4.6 ME Box plot Gradient Boosting Regressor . . . . . . . . . . . . . . . 21
4.7 Accuracy Box plot for Support Vector Regressor . . . . . . . . . . . . 22
4.8 MAE Box plot for Support Vector Regressor . . . . . . . . . . . . . . 22
4.9 ME Box plot for Support Vector Regressor . . . . . . . . . . . . . . . 23
4.10 Accuracy Box plot for Random Forest Regressor . . . . . . . . . . . . 23
4.11 MAE Box plot for Random Forest Regressor . . . . . . . . . . . . . . 24
4.12 ME Box plot for Random Forest Regressor . . . . . . . . . . . . . . . 24
4.13 Feature Importance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5.1 Average Accuracy score plot . . . . . . . . . . . . . . . . . . . . . . . 27

5.2 Average Mean Absolute Error plot . . . . . . . . . . . . . . . . . . . . 28
5.3 Average Max Error plot . . . . . . . . . . . . . . . . . . . . . . . . . 29

A.1 Item outlet sales Vs Item weight . . . . . . . . . . . . . . . . . . . . . 36

A.2 Item outlet sales Vs Item Visibility . . . . . . . . . . . . . . . . . . . 37
A.3 Item outlet sales Vs Item MRP . . . . . . . . . . . . . . . . . . . . . 38
A.4 Item Type Vs Median Item Outlet Sales . . . . . . . . . . . . . . . . 39
A.5 Outlet Establishment Year Vs Median Item Outlet Sales . . . . . . . 40
A.6 Outlet Size Vs Median Item Outlet Sales . . . . . . . . . . . . . . . . 41
A.7 Outlet Location Type Vs Median Item Outlet Sales . . . . . . . . . . 42

v
 List of Tables

4.1 Comparison of Evaluation Results . . . . . . . . . . . . . . . . . . . . 25

vi
Chapter 1
Introduction

Earlier companies used to produce goods without considering the number of sales
and demand. For any manufacturer to determine whether to increase or decrease the
production of several units, data regarding the demand for products on the market
is required. Companies can face losses if they fail to consider these values while
competing on the market. Different companies choose specific criteria to determine
their demand and sales [1].
In today’s highly competitive environment and ever-changing consumer land-
scape, accurate and timely forecasting of future revenue, also known as revenue
forecasting, or sales forecasting, can offer valuable insight to companies engaged in
the manufacture, distribution or retail of goods[2]. Short-term forecasts primarily
help with production planning and stock management, while long-term forecasts can
deal with business growth and decision-making[1].
Sales forecasting is particularly important in the industries because of the limited
shelf-life of many of the goods, which leads to a loss of income in both shortage and
surplus situations. Too many orders lead to a shortage of products and still too
few orders lead to a lack of opportunity. Therefore, competition in the food market
is continuously fluctuating due to factors such as pricing, advertisement, increasing
demand from the customers[3].
Managers usually make sales predictions randomly. Professional managers, how-
ever, become hard to find and not always available (e.g., they can get sick or leave).
Sales predictions can be assisted by computer systems that can play the qualified
managers’ role when they are not available or allow them to make the right decision
by providing potential sales predictions. One way of implementing such a method is
to try and model the professional managers’ skills inside a computer program[4].
Alternatively, the abundance of sales data and related information can be used
through Machine Learning techniques to automatically develop accurate sales pre-
dictive models. This approach is much simpler. It is not prejudiced by a single sales
manager’s particularities and is flexible, which means it can adapt to data changes.
It has, however, the potential to overestimate the accuracy of the prediction of a
human expert, which is normally incomplete. For example, once companies used
to produce the products without taking into consideration the number of sales and
demand as they faced several problems. Since they don’t know how much to sell, for
any manufacturer to decide whether to increase or decrease the number of units, data
regarding the consumer demand for products is essential. If companies do not con-
sider these principles when competing in the market, they will face losses. Different
companies choose different parameters to determine their market and sales.

1
Chapter 1. Introduction 2

There are several ways of forecasting sales in which companies have previously
focused on various statistical models such as time series and linear regression, feature
engineering and random forest models to obtain future sales and demand prediction.
Time series contains data points that are stored over a fixed period and are used
to forecast the future. Time series is a collection of data points which are collected
in period at sequential, evenly spaced points. The most important components to
analyze are patterns, seasonality, irregularity, cyclicity.
Linear regression is a mathematical tool used to forecast past values. It can help
to determine the underlying trends and address cases involving overstated rates[5][6].
Feature engineering is the use of data on domain knowledge and the development
of features to make predictive Machine Learning models more accurate. It makes
for deeper data analysis and a more useful perspective[7]. A decision tree is a fun-
damental principle behind a model of random forests. The decision tree approach
is a technique used in data mining to forecast and classify data. The decision tree
approach does not provide any conceptual understanding of the issue itself. Random
forest is the more sophisticated method that allows and merges many trees to make
decisions. The random forest model results in more accurate forecasts by taking out
an average of all individual tree decision predictions.
The entire data set is usually divided into two parts, namely the training data and
the test data. Training data is a data that is used to train the model, and test data
is the data used to evaluate the trained model. A classical approach is 80-20 split,
stating that 80 percent of the data is used to train the model, and the remaining 20
percent of the data is used to test the model. But approaches like stratified K-fold
cross-validation are known to provide good results. There were many cross-validation
variants, such as simple k-folds, leave one out, stratified k-fold cross-validation, and
so on[8][9].

1.0.1 Aims and Objectives

This thesis aims to develop a Machine Learning model that can predict the sales of
products from different outlets. Several objectives were drawn to attain the goal:

Objectives:
• Converting data into an appropriate form using various preprocessing tech-
niques for the implementation of Machine Learning algorithms.

• Finding critical features that will most influence sales of the product.

• To determine the appropriate Machine Learning algorithm for sales forecasting.

• Selecting various metrics to compare the performance of the applied Machine

Learning algorithms.

1.0.2 Research Questions

Two research questions have been defined for this study to accomplish the aim. They
are defined as follows:
Chapter 1. Introduction 3

RQ1:
What are the critical features that influence product sales?

Motivation:
The motivation of this research question is to find critical features in the data that
can be useful while experimenting for RQ2 to build the Machine Learning model.
This will help us reduce computational power and improves the quality of the results.

RQ2:
What is the best suitable algorithm for sales and demand prediction using Machine
Learning techniques?

Motivation:
The critical features identified from RQ1 are used to develop the Machine Learning
model using different algorithms. These models are compared by using various met-
rics such as accuracy score, mean absolute error, and max error to select the best fit
model for the data.

1.1 Background
There are several methods for forecasting future demand for the goods and services
a business provides. The forecasts are used for planning production and business
activities, purchasing materials, inventory management, scheduling work hours, ad-
vertising, and often more across most industries. Traditional forecasting approaches
were primarily focused on experienced employee opinions or statistical analysis of
past data, but in recent years Machine Learning techniques have been implemented
with great success in this field.

1.1.1 Data Mining

Data mining is described as a process for extracting usable data from a larger col-
lection of raw data using statistical, artificial intelligence, Machine Learning and
pattern recognition methods[10][11]. Data Mining is increasingly seen as a step in
a systematic and iterative process of knowledge discovery, in which automated pat-
tern recognition methods are combined with expert knowledge of the analyst. This
process is called the Knowledge Discovery in Databases (KDD) process[12].

1.1.2 Machine Learning

Machine Learning is the area of study which enables machines to learn without being
explicitly programmed[13]. Machine Learning is defined as the computer program
learns from experience E with respect to some class of tasks T and performance
measure P when its performance at tasks in T, as measured by P, strengthens with
Chapter 1. Introduction 4

Figure 1.1: Types of Machine Learning

experience E[14]. In general, Machine Learning is a program that can manage various
tasks by analyzing and exploring data[15].
Common Machine Learning applications such as email spam detection, credit card
fraud, stock predictions, smart assistants, product recommendations, self-driving
cars, sentiment analysis, etc.

Supervised Learning:
The most popular model for performing Machine Learning processes is supervised
learning. It is commonly used for data where the mapping between input-output data
is accurate. Supervised learning is the subset of Machine Learning which concentrates
on learning a model of classification or regression, that is, learning from labeled test
data[15].

Unsupervised Learning:
The data is not explicitly labeled into different classes in the case of unsupervised
learning that is there is only unlabeled data. By identifying implicit patterns the
model can learn from the data. Unsupervised Learning categorizes the densities,
structures, related segments, and other similar properties based on the data[16].

Reinforcement Learning:
Reinforcement Learning is a sub-field of Machine Learning. In a given scenario,
it is about taking appropriate action to optimize reward. Various algorithms and
computers are employed to determine the best possible action or path it will follow
in a specific scenario. Reinforcement learning varies with supervised learning in such
a way that the training data has the answer key with it in supervised learning such
that the model is trained with the correct response itself while in reinforcement
learning there will be no response but the reinforcement agent determines how to
Chapter 1. Introduction 5

execute the task. It is required to learn from its experience, in the absence of training
data[15].

1.1.3 Machine Learning Algorithms:

Forecasting means predicting events of the future, typically based on previous records.
For a long time, statistical models were commonly used for the conducting of pre-
dictions. The role of generalization in Machine Learning has been considered. In the
case when a new product or store is introduced, this effect could be used to make
sales predictions because there is a limited amount of historical data for a particular
time series[17]. In this thesis we have used supervised learning algorithms such as
Support Vector Machines, Random Forest Regression, Gradient Boosting, and Sim-
ple Linear Regression. These can make it easier to find better outcomes compared
to traditional analytical techniques of time series[18].

Simple Linear Regression

Simple linear regression is useful for defining a relationship between two continuous
variables. One is an indicator or independent variable and another is an answer or
dependent variable. It looks for a statistical relationship, but not a deterministic
one. The relationship between the two variables is said to be deterministic if one
variable can be precisely represented by the other[19]. For example, it is possible to
correctly forecast Fahrenheit by using temperature in degree Celsius. The mathemat-
ical equation is not sufficient to assess the association between the two variables. For
example, the relationship between weight and height. The Equation for the Simple
linear regression is:

Y = a + bX
where Y is the expected value of the dependent variable y for every specified
value of the independent variable X, a is the intercept, b is the regression coefficient
and X is independent variable.

Gradient boosting Regression

Gradient boosting is some kind of enhancement in Machine Learning. It is based
on the premise that, when combined with previous ones, the best possible current
iteration will minimize the maximum prediction error. The key idea for this next
iteration is to set the target outcomes to minimize the error.
One of the most successful Machine Learning models for predictive analytics is
the Gradient Boosted Regression Trees (GBRT) model (also called Gradient Boosted
Machine or GBM), which makes it an industrial workhorse for Machine Learning.
The Boosted Trees Model is a type of additive model that combines decisions from a
sequence of base models to make predictions[20]. One can write this class of models
more formally, as:

g(x) = f 0(x) + f 1(x) + f 2(x) + f 3(x) + .....

Chapter 1. Introduction 6

Where the final classifier g is the amount of the specific classifiers fi. Each base
classifier is a simple decision tree for model boosted trees. This broad approach of
using multiple models is called model ensembling to achieve better predictive perfor-
mance. Unlike Random Forest, which independently builds all the base classifiers,
each using a subsample of data, GBRT uses a particular technique of assembly called
gradient boosting[20].

Support Vector Machine

Support Vector Machine or SVM is one of the most common Supervised Learning
algorithms used for both Classification and Regression issues. The SVM algorithm
aims to build the best line or decision boundary that can divide n-dimensional space
into conveniently place the new data point in the right category in the future. The
optimal choice boundary is called a hyper plane. SVM chooses extreme points vectors
that help to create a hyper plane. Such extreme cases are called help vectors[21].
The equation for Support Vector Regression is:

f (x) = x β + b

Random Forest Regression

Random Forest is one of the most powerful Machine Learning frameworks for pre-
dictive analytics. A random forest method is a type of discrete structure that allows
predictions by integrating decisions from a series of simple models[22]. More formally,
this subset of models can be written as:

g(x) = f 0(x) + f 1(x) + f 2(x) + f 3(x) + .....

Where the initial configuration g is the number of the initial specific model’s fi.
Here, any base classifier is a simple decision tree. This wide-ranging technique of
using multiple models to improve predictive performance is called model assembling.
In random woods, all baseline models are built independently using a separate subset
of results.

1.1.4 Selection of Machine Learning Algorithms

For every problem, choosing an algorithm is not a trivial decision. There is no proper
algorithm that works for any problem, but few algorithms are widely recognized for
performing the algorithms better than others in some cases. One can not assume
the more accuracy from the algorithms for all types of data, accuracy will differ from
data to data. In this thesis Machine Learning Algorithms such as Simple Linear
Regression, Gradient Boosting Regression, Support Vector Regression, and Random
Forest Regression were considered in which they expected to perform well on the
issues.
Chapter 1. Introduction 7

1.1.5 Selection of Performance Metrics

To identify the appropriate algorithm one needs to evaluate the results and then
we can predict it. In this case, the accuracy score would play a crucial role while
measuring the performance of the algorithm. For calculating the average magnitude
of errors mean absolute error metric will be used in this study. In real-time data
there might be a chance of worst-case error between the actual value and predicted
value in this particular scenario max error is used.
Chapter 2
Related Work

Previously a lot of sales and demand forecasting work was performed using Machine
Learning. Most of the work in this research will concentrate on the sales of food
items.
Due to the importance of forecasting in various fields, there are so many different
types of approaches taken previously, some of the methods such as Machine Learn-
ing models, hybrid models, and statistical models. To handle this work, some of
the statistical methods such as auto regressive moving average (ARMA) and auto
regressive integrated moving average (ARIMA) will be helpful[4].
İrem İşlek and Şule Gündüz Öğüdücü experimented with the use of bipartisan
graphic clusters that clustered different warehouses according to the sales behavior.
They addressed the application by applying the Bayesian network algorithm in which
they managed to produce the enhanced forecasting experience[23].
Grigorios Tsoumakas had used Machine Learning techniques to perform a survey
on the forecasting of food sales. They had addressed data analyst design decisions
such as temporal granularity, output variable, and input variables in this survey[4].
In this paper the authors experimented by taking the point of sale (POS) as internal
data and even external data by considering different environments to enhance the
efficiency of demand forecasting. They considered different Machine Learning algo-
rithms such as Boosted Decision Tree Regression, Bayesian Linear Regression, and
Decision Forest Regression for evaluation[24].
The paper’s authors had researched interestingly about customers coming to the
restaurants using Random Forests, k-nearest neighbor, and XGBoost. They chose
two real-world data sets from different booking sites and also made different input
variables from restaurant features. The results have shown that XGBoost is the most
appropriate model for the dataset[25].
Holmberg and Halldén had observed that regular restaurant sales to be influenced
by the weather. They considered two Machine Learning algorithms as XGBoost
and neural network, and the results showed that the XGBoost algorithm is more
accurate than the other algorithm, and they also found that they had improved
their model performance by 2-4 percentage points by taking weather factors into
consideration. To improve accuracy, they had considered numerous variables such as
date characteristics, sales history, and weather factors[26].
Most of the recent studies focused on sales modeling without considering the
relationship between the training and testing data, they used training data directly.
This causes many errors which lead to a reduction in accuracy. Recent studies
have suggested clustering techniques to separate the entire forecasting data into

8
Chapter 2. Related Work 9

several clusters of predictable data before designing predictable models to minimize

computational time and achieve effective evaluating performance[27].
In particular, Support Vector Machine(SVM) had been applied to demand fore-
casting. Garcia et al. (2012), in their study, proposed an intelligent model that relies
on supporting vector machines to deal with issues relating to the allocation and rev-
elation of new models. Kandananond (2012) showed that SVM surpassed Artificial
Neural Networks in estimating demand for consumer goods[28].
Previously, most of the studies focused on considering the metrics as mean ab-
solute error, mean squared error, median absolute error, and k-fold cross validation
is used for training and testing data. Metrics like max error, accuracy, and mean
absolute error are considered in this research. In this study stratified K-fold cross-
validation technique is used for training and testing to increase the efficiency of the
results. In this study a suitable algorithm is chosen for sales forecasting.
Chapter 3
Method

In this thesis research questions are answered by using research methods. Research
aspects for this work are examined by the execution of the experiments.

3.1 Experiment
An experiment is chosen for the first research question i.e. correlation. Each data
attribute can be selected by applying feature selection methods like data correlation
and which will make the predictable attributes more accurate. This will reduce a
lot of strain on the Machine Learning model during pre-processing and cleansing
the data. For the second research question an experiment is chosen because the
experiments provide control over factors and a deeper understanding of many com-
mon research techniques such as a case study or survey[29]. One can describe the
procedure followed in this experiment as follows:

• Extracting the data required for the sales.

• Applying specified Machine Learning (supervised) algorithms.

• The performance of the output can be enhanced by comparing metrics such as

accuracy score, mean absolute error and max error.

• Based on assessment tests, the best suitable algorithm can be selected.

3.2 Experimentation Environment

Python
Python is a commonly used high-level programming language, it was designed by
Guido van Rossum which can be easy to interpret and read[30]. Python has specific
functionality and is convenient to be used for both quantitative and analytical com-
putational purposes. Data Science Python is popularly used and, as well as being a
dynamic and open source language, is a top choice. Its massive libraries are also used
to manipulate the data however for a beginner data analyst they are really simple
to learn[31]. The python libraries used in this thesis are briefly described as follows:

10
Chapter 3. Method 11

NumPy
NumPy is a library that consists of multidimensional array objects and a set of array
processing routines. NumPy is used along with SciPy and Matplotlib packages. This
combination is used for technical computing. Mathematical and logical operations
are performed with the help of NumPy[32].

Pandas
Pandas is a software library that is designed for manipulating the data and analysis
in a python programming language. It is open-source which is released under the
BSD license of three clauses. It is based on the Numpy package, and the DataFrame
is its main data structure[33].

Matplotlib
Matplotlib is a module of Python used to plot the attractive Graphs. Visual rep-
resentation in data science is a significant step. One can quickly understand how
data is split by using visual representation. There are many libraries to represent
the data, but the matplotlib is very widely known and easier to visualize[34].

SKlearn
Scikit-learn is a free python library. It features multiple clustering classification
and regression algorithms including random forests, DBSCAN, k-means, gradient
boosting, support vector machines, and gradient boosting which is programmed to
interface with the NumPy and SciPy libraries[35].

Seaborn
Seaborn is a open-source python library that is used for statistical graphics. It offers
a data set-oriented API to analyze relationships among different variables, as well as
resources to select color palettes that truly in the data[36].

3.3 Data overview

In this thesis, there is labeled sales data from different items from different outlets
that provide information such as item type, item price, outlet type, etc. These data
were extracted from various sources and will be used to train and improve the model
for Machine Learning. In the dataset being analyzed there are 8523 instances and
12 attributes. The dataset has been properly divided into training and testing data
that can be described in the sections below.

3.4 Feature Selection

There are various types of factors that can make the model of Machine Learning more
effective on any given task. One of the methods of feature selection is data correlation
Chapter 3. Method 12

Figure 3.1: Dataset description

Figure 3.2: Dataset summary

Chapter 3. Method 13

Figure 3.3: Correlation values

which will have a major impact on the model’s performance. This will reduce a lot of
strain on the Machine Learning model during preprocessing and cleansing the data.
The data attributes chosen for training the Machine Learning model would have
a major impact on the efficiency of the model. Because of the irrelevant features
that are presented, the model output will be reduced. The feature selection method
provides an efficient way to remove data redundancy and irrelevant data that helps
to reduce computation time, improve accuracy, and also enhance understanding of
the model[37].
The selection of features plays a crucial role in classification and involves selecting
a subset of features that reflect the complete attributes that currently exist. Feature
selection techniques are intended to improve classification efficiency by selecting the
essential features from the data sets according to particular algorithms.

3.4.1 Data Correlation Method

Data correlation is a method that helps to predict one attribute from another at-
tribute and is used as a basic quantity in many modeling techniques. If one feature
increases, the correlation will be positive, so the other feature increases as well and
negative if one feature increases there will be a reduction in another. If there is no
relation between any two attributes then it is said to be no correlation[38]. If there
is a linear relationship between the constant variables then the Pearson correlation
coefficient is used. If there is a non-linear relation between the constant variables
then the Spearman correlation coefficient is used.
Since the considered data set is linear so the Pearson correlation coefficient is
used for the selection of features in this study. This correlation for all the attributes
is shown in figure 3.4. To improve the efficiency of the Machine Learning model, the
attributes that have negative correlations were removed. It is a statistic measuring
the linear correlation of two variables X and Y. It has a value between +1 and 1,
where 1 is a linear positive correlation, 0 is not a linear correlation and 1 is a linear
negative correlation[39].
The motivation for considering the correlation is when people know a score on
one measure, they can make a prediction of another measure that is highly related
to it more accurate. The more accurate the prediction, the stronger the relationship
between the variables.
Chapter 3. Method 14

The heat map for correlation between non-numerical attributes is plotted as follows:

Figure 3.4: Heat map

Chapter 3. Method 15

3.5 Feature Importance

Feature Importance refers to a class of approaches for assigning values to input
features to a predictive model which determines the relative significance of each
factor while forecasting[40].
Feature importance scores provide overview into the model. Most significant
scores are determined using a prediction approach that was fitted to the dataset.
Inspecting the score of importance gives insight into that particular model and what
features are the most essential and least important to the model while making a
prediction. This is a type of interpretation of the model that can be carried out for
those models that encourage it.
Feature Importance can be used to enhance a predictive model. This can be
accomplished by selecting those features to remove (lowest scores) or those features
to retain, using the importance scores. This is a type of selection of features, and
can simplify the modeling problem, accelerate the modeling process, and in certain
cases improve model performance[40].

3.6 Data preprocessing

Before applying Machine Learning algorithms some of the missing values have been
found which can impact the model’s output so this should be handled. The ’item
weight’ and ’outlet size’ attributes have 17 percent, and there is 28 percent of missing
values. To make the dataset more efficient, these missing values will be replaced by
the most promising values. There’s more correlation between two of the different
attributes with similar work. Removing one of the attributes will make the work
better. The redundant values such as LF and reg provided in the attribute of item
fat content will be treated and these redundant values will be replaced accordingly.
The least value for an ’item visibility’ attribute is zero which makes no sense for the
dataset.

3.6.1 Encoding Categorical Values

Categorical data contains label values that are considered nominal values. Each
value has categories of different types. Besides, a few of the groups have a normal
relationship with each other is known as natural ordering. The categorical data can
be converted into numerical data to improve the efficiency of the Machine Learning
model[41].

One Hot Encoding

One hot encoding is the method where the data is represented in binary format and
included as a feature. It is one of the most common methods, comparing each level
of the numerical variable with a fixed starting point. In this thesis for the data set
that had taken, one hot encoding is used to represent categorical variables as binary
vectors[41].
This approach results in a dummy variables trap because it is easy to predict the
outcome of one variable with the support of the existing variables. This trap leads
Chapter 3. Method 16

Figure 3.5: Before One Hot Encoding

Figure 3.6: After One Hot Encoding

Chapter 3. Method 17

to a multicollinearity problem. It occurs when the independent features become

dependent upon each other. To overcome the multicollinearity problem one of the
dummy variables should be dropped. The following figure represents the before and
after one hot coding.

3.6.2 Stratified K-fold Cross-Validation

Cross-validation (CV) is a procedure of statistical analysis used to assess the effec-
tiveness of a Machine Learning technique, as well as a re-sampling method used to
validate an algorithm if there is insufficient data[42].
Stratification is the process of rearranging the data to ensure each fold is a good
representative of the whole. Data splitting into folds may be controlled by criteria
such as ensuring where each fold has the same ratio of outcomes with a given cate-
gorical value, such as the class outcome value. This process is called stratified k-fold
cross-validation[43]. Common techniques of cross-validation include K-fold cross-
validation, Stratified K-fold cross-validation, and cross-validation leave-one-out.
The motivation behind the 10-fold stratified cross-validation is that the estimator
has a lower variance than a single hold-out set estimation method which could be
very essential if there is a limited amount of data. There will be plenty of variance in
the results estimate for various data samples, or for specific data partitions to create
training and test sets. The 10-fold stratified cross-validation removes this variance
by comparing more than 10 separate partitions, thereby making the performance
estimate less sensitive to data partitioning.

3.7 Performance Metrics

Several metrics can be used while evaluating how well a model is performing. It is
necessary to understand how each metric measures to select the evaluation metric to
better assess the model. This thesis main objective was to compare the performance
of Machine Learning techniques by evaluating all of these performance metrics such
as Accuracy score, Mean Absolute Error, and Max error.

3.7.1 Accuracy score

Accuracy is known as the ratio of a several correct predictions(both true positives
and true negatives) to the total number of data points[44].
FN + FP
AccuracyScore =
N
Where FN is false negative, FP is false positive and N is total number of predic-
tions.

3.7.2 Max Error

The function max error measures the maximum standard errors, a metric represent-
ing a worse-case error between the expected value and the actual value. Max error
Chapter 3. Method 18

would be 0 on the test set in a properly fitted single-output regression analysis, and
while this would be extremely impossible in the modern world, this measurement
indicates the amount of error the model has when it was placed in[45].

M axError(y, x) = max(|yi − xi |)
Where Yi describes the actual values, Xi describes the expected values.

3.7.3 Mean Absolute Error

Mean Absolute Error is a process performance measure that is used for regression
models. A model’s mean absolute error concerning a test data set is the average of
the actual values on all instances in the test set of the specific prediction errors. For
instance, every predictive error is the difference between the predicted value and the
actual value. Mean Absolute Error is one of several metrics for summing up and
measuring the Machine Learning model’s performance[46].
nsamples −1
1 
M AE(y, x) = ( ) |yi − xi |
nsamples i=0

Where Yi describes the actual values, Xi describes the expected values.

Chapter 4
Results

Simple Linear Regressor, Gradient Boosting Regressor, Random Forest Regressor and
Support Vector Regressor are trained with the set of data using a 10-fold stratified
cross-validation approach that dynamically selected the training and testing with
fixed proportion each time and the efficiency was calculated using max error, mean
absolute error and accuracy metrics.

4.1 Simple Linear Regressor

The box plot in Figure 4.1 shows the accuracy score(ACC) obtained by the Simple
Linear Regressor during a 10-fold stratified cross-validation test. The upper box-plot
represents a maximum accuracy of 84.099 percent, the middle quartile represents a
median accuracy of 81.2 percent, and the lower quartile of the box-plot represents a
minimum accuracy of 73.95 percent.
The box plot in Figure 4.1 shows the Mean Absolute Error(MAE) obtained by
Simple Linear Regressor during a 10-fold stratified cross-validation test. The upper
box plot represents the maximum MAE of 4.6773, the middle quartile represents a
median MAE of 3.2127 and the lower quartile of the box plot represents a minimum
MAE of 2.5208.
The box plot in Figure 4.1 shows the Max Error(ME) obtained by Simple Linear
Regressor during a 10-fold stratified cross-validation test. The upper box plot rep-
resents the maximum ME of 0.5118, the middle quartile represents a median ME of
0.4917 and the lower quartile of the box plot represents a minimum ME of 0.4731.

4.2 Gradient Boosting Regressor

The box plot in Figure 4.4 shows the accuracy score(ACC) obtained by Gradient
Boosting Regressor during a 10-fold stratified cross-validation test. The upper box
plot represents the maximum accuracy score of 91.2 percent, the middle quartile
represents a median accuracy score of 86.27 percent and the lower quartile of the
box plot represents a minimum accuracy score of 78.3 percent.
The box plot in Figure 4.5 shows the Mean Absolute Error(MAE) obtained by
Gradient Boosting Regressor during a 10-fold stratified cross-validation test. The
upper box plot represents the maximum MAE of 4.43, the middle quartile represents
a median MAE of 3.15 and the lower quartile of the box plot represents a minimum
MAE of 2.77.

19
Chapter 4. Results 20

The outcomes that Simple Linear Regressor obtains are as follows:

Figure 4.1: Accuracy Box plot for Simple Linear Regressor

Figure 4.2: MAE Box plot for Simple Linear Regressor

Figure 4.3: ME Box plot for Simple Linear Regressor

Chapter 4. Results 21

The outcomes that Gradient Boosting Regressor obtains are as follows:

Figure 4.4: Accuracy Box plot for Gradient Boosting Regressor

Figure 4.5: MAE Box plot Gradient Boosting Regressor

Figure 4.6: ME Box plot Gradient Boosting Regressor

Chapter 4. Results 22

The outcomes that Support Vector Regressor obtains are as follows:

Figure 4.7: Accuracy Box plot for Support Vector Regressor

Figure 4.8: MAE Box plot for Support Vector Regressor

The box plot in Figure 4.6 shows the Max Error(ME) obtained by Gradient
Boosting Regressor during a 10-fold stratified cross-validation test. The upper box
plot represents the maximum ME of 0.464, the middle quartile represents a median
ME of 0.441 and the lower quartile of the box plot represents a minimum ME of
0.425.

4.3 Support Vector Regressor

The box plot in Figure 4.7 shows the accuracy score(ACC) obtained by Support
Vector Regressor during a 10-fold stratified cross-validation test. The upper box
plot represents the maximum accuracy score of 88.45 percent, the middle quartile
represents a median accuracy score of 84.82 percent and the lower quartile of the box
plot represents a minimum accuracy score of 78.3 percent.
The box plot in Figure 4.8 shows the Mean Absolute Error(MAE) obtained by
Chapter 4. Results 23

Figure 4.9: ME Box plot for Support Vector Regressor

The outcomes that Random Forest Regressor obtains are as follows:

Figure 4.10: Accuracy Box plot for Random Forest Regressor

Support Vector Regressor during a 10-fold stratified cross-validation test. The upper
box plot represents the maximum MAE of 4.3507, the middle quartile represents a
median MAE of 3.1647 and the lower quartile of the box plot represents a minimum
MAE of 2.7238.
The box plot in Figure 4.9 shows the Max Error(ME) obtained by Support Vec-
tor Regressor during a 10-fold stratified cross-validation test. The upper box plot
represents the maximum ME of 0.4686, the middle quartile represents a median ME
of 0.4485 and the lower quartile of the box plot represents a minimum ME of 0.4343.

4.4 Random Forest Regressor

The box plot in Figure 4.10 shows the accuracy score(ACC) obtained by Random
Forest Regressor during a 10-fold stratified cross-validation test. The upper box
plot represents the maximum accuracy score of 91.35 percent, the middle quartile
Chapter 4. Results 24

Figure 4.11: MAE Box plot for Random Forest Regressor

Figure 4.12: ME Box plot for Random Forest Regressor

Chapter 4. Results 25

represents a median accuracy score of 87.72 percent and the lower quartile of the box
plot represents a minimum accuracy score of 78.31 percent.
The box plot in Figure 4.11 shows the Mean Absolute Error(MAE) obtained by
Random Forest Regressor during a 10-fold stratified cross-validation test. The upper
box plot represents the maximum MAE of 5.8058, the middle quartile represents a
median MAE of 4.458 and the lower quartile of the box plot represents a minimum
MAE of 3.4156.
The box plot in Figure 4.12 shows the Max Error(ME) obtained by Random
Forest Regressor during a 10-fold stratified cross-validation test. The upper box plot
represents the maximum ME of 0.6568, the middle quartile represents a median ME
of 0.6135 and the lower quartile of the box plot represents a minimum ME of 0.5964.

4.5 Feature Importance

Figure 4.13: Feature Importance

Figure 4.13 shows that the feature importance of price of the products would
depend primarily on the sales followed by the type of outlet and grocery store and
the rest of the features would not even close to these features. There will surely be
a huge impact on sales forecasting with these features.

4.6 Evaluation Results

Algorithms Accuracy Score Mean Absolute Error Max Error

Random Forest Regression 87.72 percent 3.15 0.441
Gradient Boosting Regression 86.27 percent 3.198 0.491
Support Vector Machine 84.82 percent 3.21 0.448
Simple Linear Regression 81.2 percent 3.212 0.491

Table 4.1: Comparison of Evaluation Results

Chapter 4. Results 26

Table 4.1 shows the comparison of evaluation results where Random Forest Re-
gression performed well with all the metrics accuracy score, Mean Absolute Error
and Max Error. Random Forest Regression had the minimum error in predicting the
sales when compared to the Simple Linear Regression, Gradient Boosting regression
and Support Vector Machine. Simple Linear Regression demonstrated the worst per-
formance with the highest error in all the metrics. A simplified tabular form based
on the results is created above.
Chapter 5
Analysis and Discussion

5.1 Comparative analysis of Performance Metrics

In this section the average determined on the basis of ten iterations that were con-
sidered for all the metrics.

5.1.1 Average Accuracy Score

Figure 5.1: Average Accuracy score plot

Figure 5.1 shows the average accuracy score of the 10-fold stratified cross-validation
obtained by the Simple Linear Regressor is 81.2 percent, followed by the Gradient
Boosting Regressor with 86.27 percent accuracy score, then SVR with the 84.82 per-
cent accuracy score and finally Random Forest Regressor with 87.72 percent accuracy

27
Chapter 5. Analysis and Discussion 28

score. From figure 5.1, it can be seen that Random Forest Regressor is the best per-
former with approximately 88 percent accuracy score compared to other methods,
and Simple Linear Regressor is the poor performer with an accuracy score of 81.2
percent.

5.1.2 Average Mean Absolute Error

Figure 5.2: Average Mean Absolute Error plot

Figure 5.2 shows the average MAE of the 10-fold stratified cross-validation ob-
tained by Simple Linear Regressor is 3.21 error, followed by Gradient Boosting Re-
gressor with 3.19 error, then SVR with 3.21 error, and finally Random Forest Re-
gressor with 3.15 error respectively. From figure 5.2, it can be shown that Random
Forest Regressor is the best performer with less error relative to other approaches
and with the highest error, Simple Linear Regressor is the poor performer.

5.1.3 Average Max Error

Figure 5.3 shows the average MAE of the 10-fold stratified cross-validation obtained
by Simple Linear Regressor is 0.49 error, followed by Gradient Boosting Regressor
with 0.491 error, then SVR with 0.448 error, and finally Random Forest Regressor
with 0.441 error respectively. From figure 5.3, it can be shown that Random Forest
Regressor is the best performer with less error relative to other approaches and with
the highest error, Simple Linear Regressor is the poor performer.
Chapter 5. Analysis and Discussion 29

Figure 5.3: Average Max Error plot

5.2 Discussion
RQ1: What are the critical features that influence product sales?

Answer:
It was clearly observed in figure 4.13 that the price of the products and followed by
the type of outlet and grocery store will heavily influence the product sales.

RQ2:
What is the best suitable algorithm for sales and demand prediction using Machine
Learning techniques?

Answer:
Random Forest Regression is the most appropriate algorithm for forecasting the
product sales. When compared to the Simple Linear Regression, Gradient Boosting
Regression and Support Vector Regression, Random Forest Regression technique will
produce the least error while predicting the product sales.
Average accuracy score, mean absolute error and max error for the Random Forest
Regressor across the 10-fold stratified cross-validation is 87.72 percent, 3.15 and 0.44
error respectively which is quite impressive compared to other techniques. Simple
linear Regressor produced the very poor results compared to the other techniques,
Chapter 5. Analysis and Discussion 30

the average accuracy score, mean absolute error and max error across the 10-fold
stratified cross-validation is 81.2 percent, 3.21, 0.49 error which is the poor one. And
we can also observe from figure 4.13, Item price and outlet type grocery store are
the critical features that will mainly influence the product sales. If the sales forecast
is carried on every day across a large number of stores speed will play a key aspect
in this process. Another useful metric to train the model, which will also play a
crucial role while training several algorithms. The other important measure is the
time required to train the model, which will also play a critical role while training
different types of algorithms.

5.3 Contributions
As there are so many ongoing experiments that use statistical approaches and some
traditional methods to focus on predicting item sales. Most researches have experi-
mented by taking a single algorithm to predict sales. In this thesis Machine Learning
algorithms such as Simple Linear Regression, Support Vector Regression, Gradient
Boosting algorithm, and Random Forest Regression are considered for prediction and
the most effective metrics such as accuracy, mean absolute error, and max error are
considered for measuring algorithm efficiency. This method will be very beneficial in
the future for advanced item sales forecasting.

5.4 Validity Threats

5.4.1 Internal Validity
Proper preprocessing of data will be done. There may be a high chance of less
accuracy if one can not perform data preprocessing properly.

5.4.2 External Validity

The data and findings used for the experiments are justified and relevant. This
problem can be solved if the performance metrics and algorithms are not chosen
properly. Proper selection of algorithms and performance metrics will be done.
Chapter 6
Conclusions and Future Work

6.1 Conclusion
Sales forecasting plays a vital role in the business sector in every field. With the
help of the sales forecasts, sales revenue analysis will help to get the details needed
to estimate both the revenue and the income. Different types of Machine Learning
techniques such as Support Vector Regression, Gradient Boosting Regression, Simple
Linear Regression, and Random Forest Regression have been evaluated on food sales
data to find the critical factors that influence sales to provide a solution for forecasting
sales. After performing metrics such as accuracy, mean absolute error, and max error,
the Random Forest Regression is found to be the appropriate algorithm according
to the collected data and thus fulfilling the aim of this thesis.

6.2 Future Work

In future work one can attempt performance metrics such as time while predicting
the sales. These metrics can play a crucial role in evaluating multiple Machine
Learning algorithms. And also one can attempt to implement more accurate data
in the continued study. Machine Learning has the advantage of analyzing data and
key variables so that you can aim to develop a systematic approach using a variety
of Machine Learning techniques.

31
 References

[1] Patrick Bajari, Denis Nekipelov, Stephen P Ryan, and Miaoyu Yang. Ma-
chine learning methods for demand estimation. American Economic Review,
105(5):481–85, 2015.

[2] Kris Johnson Ferreira, Bin Hong Alex Lee, and David Simchi-Levi. Analytics for
an online retailer: Demand forecasting and price optimization. Manufacturing
& Service Operations Management, 18(1):69–88, 2016.

[3] Ankur Jain, Manghat Nitish Menon, and Saurabh Chandra. Sales forecasting
for retail chains, 2015.

[4] Grigorios Tsoumakas. A survey of machine learning techniques for food sales
prediction. Artificial Intelligence Review, 52(1):441–447, 2019.

[5] Xiaogang Su, Xin Yan, and Chih-Ling Tsai. Linear regression. Wiley Interdis-
ciplinary Reviews: Computational Statistics, 4(3):275–294, 2012.

[6] Toby J Mitchell and John J Beauchamp. Bayesian variable selection in linear
regression. Journal of the american statistical association, 83(404):1023–1032,
1988.

[7] Zheng Li, Xianfeng Ma, and Hongliang Xin. Feature engineering of machine-
learning chemisorption models for catalyst design. Catalysis today, 280:232–238,
2017.

[8] Xinchuan Zeng and Tony R Martinez. Distribution-balanced stratified cross-

validation for accuracy estimation. Journal of Experimental & Theoretical Ar-
tificial Intelligence, 12(1):1–12, 2000.

[9] Konstantinos Sechidis, Grigorios Tsoumakas, and Ioannis Vlahavas. On the

stratification of multi-label data. In Joint European Conference on Machine
Learning and Knowledge Discovery in Databases, pages 145–158. Springer, 2011.

[10] Chris Rygielski, Jyun-Cheng Wang, and David C Yen. Data mining techniques
for customer relationship management. Technology in society, 24(4):483–502,
2002.

[11] Krzysztof J Cios, Witold Pedrycz, Roman W Swiniarski, and Lukasz Andrzej
Kurgan. Data mining: a knowledge discovery approach. Springer Science &
Business Media, 2007.

32
References 33

[12] Maike Krause-Traudes, Simon Scheider, Stefan Rüping, and Harald Meßner.
Spatial data mining for retail sales forecasting. In 11th AGILE International
Conference on Geographic Information Science, pages 1–11, 2008.
[13] Stephen Marsland. Machine learning: an algorithmic perspective. CRC press,
2015.
[14] ML documentation. https://www.mathworks.com/discovery/
machine-learning.html). Accessed: 2020-04-22.
[15] Ethem Alpaydin. Introduction to machine learning. MIT press, 2020.
[16] Arvin Wen Tsui, Yu-Hsiang Chuang, and Hao-Hua Chu. Unsupervised learning
for solving rss hardware variance problem in wifi localization. Mobile Networks
and Applications, 14(5):677–691, 2009.
[17] Bohdan M Pavlyshenko. Machine-learning models for sales time series forecast-
ing. Data, 4(1):15, 2019.
[18] Taiwo Oladipupo Ayodele. Types of machine learning algorithms. New advances
in machine learning, pages 19–48, 2010.
[19] Sanford Weisberg. Applied linear regression, volume 528. John Wiley & Sons,
2005.
[20] Gradient Boosting documentation. https://turi.com/learn/userguide/
supervised-learning/boosted_trees_regression.html). Accessed: 2020-
05-19.
[21] JN Hu, JJ Hu, HB Lin, XP Li, CL Jiang, XH Qiu, and WS Li. State-of-charge
estimation for battery management system using optimized support vector ma-
chine for regression. Journal of Power Sources, 269:682–693, 2014.
[22] Wangchao Lou, Xiaoqing Wang, Fan Chen, Yixiao Chen, Bo Jiang, and Hua
Zhang. Sequence based prediction of dna-binding proteins based on hybrid
feature selection using random forest and gaussian naive bayes. PloS one, 9(1),
2014.
[23] İrem İşlek and Şule Gündüz Öğüdücü. A retail demand forecasting model based
on data mining techniques. In 2015 IEEE 24th International Symposium on
Industrial Electronics (ISIE), pages 55–60. IEEE, 2015.
[24] Takashi Tanizaki, Tomohiro Hoshino, Takeshi Shimmura, and Takeshi Take-
naka. Demand forecasting in restaurants using machine learning and statistical
analysis. Procedia CIRP, 79:679–683, 2019.
[25] Xu Ma, Yanshan Tian, Chu Luo, and Yuehui Zhang. Predicting future visitors
of restaurants using big data. In 2018 International Conference on Machine
Learning and Cybernetics (ICMLC), volume 1, pages 269–274. IEEE, 2018.
[26] Mikael Holmberg and Pontus Halldén. Machine learning for restaurant sales
forecast, 2018.
References 34

[27] I-Fei Chen and Chi-Jie Lu. Sales forecasting by combining clustering and
machine-learning techniques for computer retailing. Neural Computing and Ap-
plications, 28(9):2633–2647, 2017.

[28] Malek Sarhani and Abdellatif El Afia. Intelligent system based support vector
regression for supply chain demand forecasting. In 2014 Second World Confer-
ence on Complex Systems (WCCS), pages 79–83. IEEE, 2014.

[29] Jason Brownlee. Introduction to time series forecasting with python: how to
prepare data and develop models to predict the future. Machine Learning Mastery,
2017.

[30] Python history. https://en.wikipedia.org/wiki/Python_(programming_

language). Accessed: 2020-04-29.

[31] Guido Van Rossum et al. Python programming language. In USENIX annual
technical conference, volume 41, page 36, 2007.

[32] Travis E Oliphant. A guide to NumPy, volume 1. Trelgol Publishing USA, 2006.

[33] Wes McKinney. Pandas, python data analysis library. see http://pandas. pydata.
org, 2015.

[34] Niyazi Ari and Makhamadsulton Ustazhanov. Matplotlib in python. In

2014 11th International Conference on Electronics, Computer and Computation
(ICECCO), pages 1–6. IEEE, 2014.

[35] Raul Garreta and Guillermo Moncecchi. Learning scikit-learn: machine learning
in python. Packt Publishing Ltd, 2013.

[36] Seaborn documentation. https://seaborn.pydata.org/introduction.

html). Accessed: 2020-04-26.

[37] Chung-Jui Tu, Li-Yeh Chuang, Jun-Yang Chang, Cheng-Hong Yang, et al. Fea-
ture selection using pso-svm. International Journal of Computer Science, 2007.

[38] Tao Zhang, Tianqing Zhu, Ping Xiong, Huan Huo, Zahir Tari, and Wanlei
Zhou. Correlated differential privacy: Feature selection in machine learning.
IEEE Transactions on Industrial Informatics, 2019.

[39] Pearson documentation. https://en.wikipedia.org/wiki/Pearson_

correlation_coefficient). Accessed: 2020-04-25.

[40] Feature Importance documentation. https://machinelearningmastery.

com/calculate-feature-importance-with-python/#:~:text=Feature%
20importance%20refers%20to%20a,feature%20when%20making%20a%
20prediction.). Accessed: 2020-06-06.

[41] Kedar Potdar, Taher S Pardawala, and Chinmay D Pai. A comparative study
of categorical variable encoding techniques for neural network classifiers. Inter-
national journal of computer applications, 175(4):7–9, 2017.
References 35

[42] Cross validation documentation. https://towardsdatascience.com/

cross-validation-explained-evaluating-estimator-performance-e51e5430ff85).
Accessed: 2020-04-28.

[43] Sebastian Raschka. Model evaluation, model selection, and algorithm selection
in machine learning. arXiv preprint arXiv:1811.12808, 2018.

[44] Accuracy documentation. https://medium.com/thalus-ai/

performance-metrics-for-classification-problems-in-machine-learning-part-i-b085
Accessed: 2020-05-01.

[45] scilearn max error. https://scikit-learn.org/stable/modules/model_

evaluation.html#max-error. Accessed: 2020-05-10.

[46] scilearn mean absolute error. https://scikit-learn.org/stable/modules/

model_evaluation.html#mean-absolute-error. Accessed: 2020-05-10.
Appendix A
Appendix

The graph below is a comparative analysis between the item sales and the item
weight.

Figure A.1: Item outlet sales Vs Item weight

36
Appendix A. Appendix 37

The graph below is a comparative analysis between the item sales and the item
visibility.

Figure A.2: Item outlet sales Vs Item Visibility

Appendix A. Appendix 38

The graph below is a comparative analysis between the item sales and the item price.

Figure A.3: Item outlet sales Vs Item MRP

Appendix A. Appendix 39

The following graph illustrates the impact of item type on item outlet sales.

Figure A.4: Item Type Vs Median Item Outlet Sales

Appendix A. Appendix 40

The following graph illustrates the impact of outlet establishment year on item outlet
sales.

Figure A.5: Outlet Establishment Year Vs Median Item Outlet Sales

Appendix A. Appendix 41

The following graph illustrates the impact of outlet sizes on item outlet sales.

Figure A.6: Outlet Size Vs Median Item Outlet Sales

Appendix A. Appendix 42

The following graph illustrates the impact of outlet location type on item outlet sales.

Figure A.7: Outlet Location Type Vs Median Item Outlet Sales

Faculty of Computing, Blekinge Institute of Technology, 371 79 Karlskrona, Sweden