Explainable AI Framework for Fraud

Detection in Financial Transactions

Vishav Pratap Singh Satyam Aniruddh kumar nagar
Department of Computer Science Department of Computer Science Department of Computer Science
Chandigarh University Chandigarh University Chandigarh University
Mohali, Punjab, India Mohali, Punjab, India Mohali, Punjab, India
[email protected] [email protected] [email protected]

\
Sejal Kapoor
Vantipalli Siddhu
Balaparameshwar Reddy Rondla Department of Computer
Department of Computer Science
Department of Computer Science Science
Chandigarh University Chandigarh University
Chandigarh University
Mohali, Punjab, India Mohali, Punjab, India
Mohali, Punjab, India
[email protected] [email protected]
[email protected]

Abstract —In the banking industry, artificial

intelligence (AI) is transforming fraud detection,
but it also brings up issues with accountability and
transparency. A lot of AI systems function as
"black boxes," which makes it challenging for I. INTRODUCTION
stakeholders to comprehend the logic underlying
In recent years, financial fraud has become an
judgments, such the reason a transaction is
reported as fraudulent. [8]Explainable Artificial increasingly critical issue for financial
Intelligence (XAI) is a collection of methods that institutions, e-commerce platforms, and
improves the overall accuracy of AI models while businesses worldwide. As the volume of
also making them more morally and financial transactions grows with the digital
comprehensibly. By giving explanations for how economy, fraudulent activities have also
AI models arrive at their predictions or
judgments, it avoids having AI systems function as
increased, resulting in significant financial
"Black boxes" and maintains predictive losses. Traditional fraud detection methods are
performance. This study investigates an often insufficient in handling large-scale data and
Explainable AI framework that combines XAI evolving fraud patterns. This has led to the
methods with deep learning and machine learning adoption of AI-powered systems that can detect
models to identify fraudulent transactions while
fraudulent transactions more accurately and
maintaining transparency. Decision Tree, Logistic
Regression, Light Gradient Boosting Machine efficiently.
(LightGBM), and Extreme Gradient Boosting
Financial fraud has grown in importance over the
(XGBoost) are the four machine learning
techniques that are covered in this research study. past few years for financial institutions, e-
Furthermore employed to provide a comparison commerce sites, and companies all over the
analysis of models are two deep learning world. The digital economy has led to a rise in
algorithms: convolutional neural networks and fraudulent activities and substantial financial
artificial neural networks. Because they are good
losses as a result of the amount of financial
at explaining model predictions, XAI techniques
like Feature Importance, Shapley Additive transactions increasing. Conventional techniques
Explanations (SHAP), and Local Interpretable for detecting fraud frequently fall short when
Model-Agnostic Explanations (LIME) are dealing with vast amounts of data and changing
employed. The study shows how integrating XAI fraud trends. As a result, artificial intelligence
methods with AI models enhances fraud detection (AI)-powered systems that can more precisely
systems' dependability, transparency, and
and effectively identify fraudulent transactions
regulatory compliance.
are being used.
Nevertheless, a lot of these AI models function
as "black boxes," which means that humans have
a difficult time understanding how they make
Keywords — Explainable AI, Fraud Detection,
decisions. Regulators, stakeholders, and end
Financial Transactions, Machine Learning,
users are concerned about the reliability and
SHAP, LIME, Transparency.
accountability of AI systems due to this lack of Dimensionality reduction: Reducing the
transparency. dimensionality of the dataset through Principal
Component Analysis (PCA) speeds up model
Fintech uses of AI, particularly in fraud detection,
training and lowers overfitting.
have shown promise in managing massive amounts
of transactional data. The interpretability of these Step 2: [3]Machine Learning and Deep Learning
models is still somewhat difficult, though. In an Algorithms
effort to increase transparency in AI judgments,
studies have looked into a variety of [4]XAI To categorize transactions as fraud or real, the
techniques, such as Shapley Additive Explanations study used two deep learning models and five
(SHAP) and Local Interpretable Model-Agnostic machine learning algorithms. These models were
Explanations (LIME). Transparency is essential in selected because to their capacity to manage
huge datasets and intricate feature connections.
the finance industry for regulatory compliance as
well as stakeholder trust. According to recent
1. Decision Tree (DT): A straightforward tree-
studies, XAI can improve financial system based model that divides the dataset according to
accountability by allowing people to spot biases feature values in order to make judgments.
and mistakes in AI models. Although they are interpretable by nature,
decision trees may overfit.
II. METHODOLOGY
2. Logistic Regression (LR): Based on a
Explainable AI (XAI) techniques[9] are used in logistic function, this linear model forecasts the
an organized manner for data preprocessing, likelihood that a transaction is fraudulent.
model training, and the integration of Although it is easily understood and
explainability techniques to make predictions straightforward, it could miss intricate patterns
comprehensible in the suggested methodology in the data.
for fraud detection. A thorough explanation of
3. Light Gradient Boosting Machine
each step is provided below:
(LightGBM): An extremely effective boosting
approach that optimizes speed and accuracy by
iteratively building trees. In particular,
LightGBM is particularly useful in handling
large-scale data.
4. Extreme Gradient Boosting (XGBoost): An
effective boosting method that builds trees one
after the other, fixing each one's mistakes. High
performance in classification tasks, such as fraud
Figure 1 — Methodology detection, is a well-known attribute of XGBoost.
Step 1: Preparing the data

The data must be cleaned and processed before

being fed into machine learning algorithms. High
Deep Learning Models:
dimensionality, uneven classes, and missing
values are common problems in fraud detection 1. Artificial Neural Networks (ANN): A multi-
datasets. The subsequent preprocessing actions layered deep learning model that can capture
are utilized: non-linear correlations in the data. With intricate
feature interactions, it works well with sizable
Class Imbalance: The dataset is probably very
datasets.
unbalanced because fraudulent transactions are
not common. Techniques like Random 2. CNNs: Originally designed for image
Oversampling—which involves replicating classification, CNNs were modified for this
minority class samples—are employed to purpose in order to identify local or spatial
overcome this. patterns in the transactional data, providing a
fresh method of fraud detection.
Making sure that all features have similar scales
is known as "feature scaling and normalization," Step 3: Model Evaluation
and it is particularly important for algorithms
that depend on feature magnitudes (like logistic Following training, the models were assessed
regression and neural networks). using the following performance measures on a
test set:
 • Accuracy: The percentage of transactions the complex model. LIME assesses if a certain
(fraudulent and genuine) that are transaction was considered fraudulent or lawful
accurately classified. using a simple surrogate model.

• Precision: the proportion of all anticipated

fraudulent transactions that were
accurately identified as fraudulent (helps III. RESULTS AND DISCUSSIONS
reduce false positives).
The main aim of this study was to create a
strong framework for detecting fraud by
• Recall (Sensitivity): The percentage of
utilizing machine learning (ML) and deep
real frauds that were accurately
learning (DL) algorithms. Additionally,
anticipated, which lowers false negatives.
Explainable AI (XAI) approaches were used to
guarantee accountability and transparency. The
• F1 Score: A balanced assessment for
study examined the outcomes of different
unbalanced datasets, calculated as the
algorithms based on performance indicators,
harmonic mean of precision and recall.
including accuracy, F1 score, AUC-ROC, and
interpretability. This allowed for an analysis of
• AUC-ROC: The Receiver Operating
the advantages and disadvantages of each
Characteristic curve's Area Under the
method. A thorough analysis of the outcomes
Curve, which gauges how well a model
from the ML and DL models can be seen
can discern between authentic and
below.
fraudulent transactions. Better
discrimination is indicated by a higher
AUC-ROC score.
Machine Learning Model Evaluation:

Step 4: XAI Integration Decision Tree model:

Using Explainable AI (XAI) approaches, the

trained models were made transparent and
addressed the "black box" character of AI models.
The XAI techniques listed below were used:

4.1 Feature Importance

Feature significance scores were retrieved for the

machine learning models in order to ascertain
which features had the most influence on the
predictions. This approach is directly provided by
tree-based models such as Random Forest,
XGBoost, and LightGBM.

4.2 Shapley Additive Explanations (SHAP) Fig. Decision Tree Confusion matrix

Logistic Regression Model:

SHAP values were calculated for each model to
take into consideration both individual and
collective forecasts. SHAP provides information
on the relative contributions of each factor to the
decision, both favorably and negatively, and has
both local and global interpretability.

4.3 [15]Local Interpretable Model-Agnostic

Explanations (LIME)

LIME was used to explain individual transactions

by using an interpretable local approximation of
 [8]
Deep Neural Networks Model Evaluation:
This section shows the findings from the analysis
of the deep neural network model that was put
into practice. The figures show the confusion
matrix for both CNN and ANN. The algorithm
fits the CNN and ANN models to the training
data for 10 epochs, with a batch size of 32 and
validation data from the test set. Next, using both
the training and test datasets, it extracts the
accuracy and loss values for each model.

Fig. Logistic Regression Confusion matrix

LightGBM Model:

Fig. ANN and CNN confusion matix

Evaluation of XAI Techniques: Several XAI

approaches, especially LightGBM and
XGBoost, were used to the top-performing
models to make sure the AI models were
comprehensible and reliable. These methods
provide insights into individual predictions
made locally as well as into the behavior of the
model globally.

Feature Importance: Tree-based models

(LightGBM, XGBoost) used feature
importance to rank features according to how
much of a contribution they made to fraud
Fig. LightGBM Confusion matrix detection. Although feature importance offered
XGBoost Model: a broad perspective, it was insufficient to
account for specific forecasts.

Fig. Feature Importance plot

[13]
Shapley Additive Explanations (SHAP):
Fig. XGBoost Confusion matrix
Model predictions were given both local and
global interpretations by SHAP values. The
contribution of each feature to the final
conclusion across all forecasts was displayed in
SHAP summary graphs.

Fig. Prediction Probabilities

Fig. SHAP Summary plot for Logistic

Regression model

Fig. Features and values for explanation

Fig. LIME Local Explanation for class fraud

The show_in_notebook method, which presents

Fig. SHAP Summary plot for LightBGM model
the explanation in a notebook format, is how the
[15]
Local Interpretable Model-Agnostic code visualizes the explanation. The feature
Explanations (LIME): LIME constructed importance values can be hidden by using the
interpretable local surrogate models around show_all = False argument.
This study's XAI techniques aid in the
individual forecasts. It was successful in
interpretation of the model's conclusions. In-
providing an explanation for the flagging of
depth, instance-specific explanations of each
some high-risk transactions by roughly
feature's role in a prediction are provided by
mimicking the original black-box model's
SHAP. LIME provides comprehensible
behavior. This method provided clear, concise explanations for individual fraud predictions by
explanations for intricate judgments, which locally approximating the model. By ensuring
increased trust in the AI system. that the AI system is transparent and accurate,
these techniques aid financial institutions in
comprehending and gaining confidence in the
fraud detection process.

The study showed that while high-performing

models such as LightGBM and XGBoost have
good fraud detection capabilities, their
interpretability is lacking. On the other hand,
while simpler models like logistic regression and V. FUTURE SCOPE
decision trees provide clarity, their predictive In order to improve accuracy and interpretability,
effectiveness suffers as a result. This trade-off is this research paper's next focus will be on
mitigated by the integration of XAI techniques, investigating more thorough integration of XAI
which balance accuracy and transparency by techniques with intricate AI architectures, such as
enabling high-performing models to justify their hybrid models that combine standard machine
predictions. learning with neural networks. Furthermore, a
useful development could be the application of
real-time adaptive learning models, which
enhance fraud detection skills over time by
IV. CONCLUSION learning from changing transaction patterns.
Federated learning holds promise for enhancing
This research emphasizes the importance of
data security and privacy in fraud detection
combining deep learning (DL) and advanced
across many financial institutions, all while
machine learning (ML) models with preserving explainability using distributed XAI
[8]
Explainable Artificial Intelligence (XAI) frameworks. Furthermore, broadening the
techniques to address the issues of application of XAI to identify biases and
interpretability and performance in financial abnormalities in the AI models itself as well as to
transaction fraud detection systems. The rising explain specific forecasts may result in more
dependence of financial institutions on artificial morally sound and open AI systems in the
intelligence (AI) solutions for fraud detection financial sector. Finally, developing AI systems
has made it imperative for these systems to that adhere to operational, ethical, and regulatory
possess transparency and accountability in order norms will require interdisciplinary cooperation
to meet regulatory requirements and foster with professionals in law, ethics, and finance.
stakeholder trust. This will guarantee the systems' broader
acceptance and confidence in crucial financial
The trade-off between accuracy and applications.
interpretability is highlighted via a comparative
examination of two DL models (ANN and
CNN) and five ML algorithms (Decision Tree,
Logistic Regression, LightGBM, and
XGBoost). Although LightGBM demonstrated
the greatest accuracy (98.3%) and AUC-ROC
score (0.96) among the models tested, more
interpretable models such as Decision Tree and
Logistic Regression demonstrated lower
efficacy in identifying intricate fraud patterns.
Even the most sophisticated models were able to
give interpretable and intelligible predictions
thanks to the application of XAI techniques like
SHAP and LIME, which added a critical layer
of transparency.

To sum up, this study shows that combining

XAI methods with high-performance AI models
provides a well-rounded approach to fraud
detection that guarantees both interpretability
and prediction accuracy. Maintaining this
equilibrium is crucial for promoting confidence,
fulfilling legal obligations, and guaranteeing
ethical AI application in the banking sector.
According to the research, explainability should
be a key component of future AI systems'
development to increase their transparency and
dependability for both technical and non-
technical users in crucial financial sectors.