Tezislər / Theses

FINANCIAL FRAUD DETECTION USING VARIOUS MACHINE 
AND DEEP LEARNING TECHNIQUES FOR BETTER 
PERFORMANCE 
Aytaj Abdullayeva 
Baku Higher Oil School 
Baku, Azerbaijan 
[e-mail] 
Supervisor: Ph.D Associate Professor Kamala Pashayeva 
Keywords: 
Fraud detection, Machine Learning, Deep Learning, Fraud detection using ML 
Tremendous number of transactions are generated everyday all over the 
world, which poses a great challenge for the banks or other financial 

---

THE 3
rd
 INTERNATIONAL SCIENTIFIC CONFERENCES OF STUDENTS AND YOUNG RESEARCHERS 
dedicated to the 99
th
anniversary of the National Leader of Azerbaijan Heydar Aliyev
110
companies to keep very secure system to prevent fraudulent and insecure 
actions. The total number of the loss caused by frauds has tripled in the last 
decade, for example, in 2021 this figure increased to more than $33 billion, 
with 6.83 cents for every $100 [1]. Therefore, according to the industry 
analysts, the cost of loss caused by the frauds will experience a huge 
increase. In order to prevent it, several rule-based systems are implemented 
by the security teams to handle the fraudulent transactions. However, this is 
not effective, as it requires more manual work and long-term processing and 
can also be outsmarted by criminals with finding various ways. Hence, one 
of the most effective and widely used method is to implement automatic fraud 
detection algorithms. In order to protect customers from the digital 
scammers, various frameworks such as Machine and Deep Learning 
techniques are widely used to identify the transaction anomalies.
Machine Learning is the combination of different computer algorithms 
and statistical modelling to perform the tasks without hard coding because 
the prediction will be made from the stored experimental knowledge. Deep 
Learning techniques is a part of ML which covers various Neural Networks 
(NN) that if trained properly, it can capture the hidden spots useful for the 
prediction of frauds [2]. Using proper methods, it is feasible to model the 
transactional behavior of each customer according to the history, so that the 
transaction can be classified. As the dataset is marked, supervised learning 
techniques with appropriate hyperparameter tuning for the unbalanced 
datasets are viable for his case. Dataset that has been used is provided by 
Vesta corporation, but some columns are masked due to the security issues. 
Different Machine and Deep Learning techniques will be passed from a grid 
to find the best method according to their metrics because the target variable 
is imbalanced. As our dataset is imbalanced, AUC (Area Under Curve) has 
been used for evaluating the model performance, it measures the area under 
the ROC curve. The higher AUC means the better distinguishability of the 
model for the positive and negative cases. When this core is near to 1, it 
means the model separability is quite good. X-axis depicts the false positive 
rate, while Y-axis presents the true positive rate of the predictions [3]. 
𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑅𝑎𝑡𝑒 =
𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑡𝑖𝑣𝑒𝑠
𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 + 𝑇𝑟𝑢𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠
, 
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑅𝑎𝑡𝑒 =
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑡𝑖𝑣𝑒𝑠
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 + 𝐹𝑎𝑙𝑠𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠
The main purpose is not to define only the fraudulent transitions but the 
fraud clients. Once the credit card owner has a fraud, their entire account 
should be labelled as a fraud. The main logic here is to define the reported 
fraudulent transactions and other transactions posterior to it. The unique 

---