Outlier Detection in Financial

Transactions Using Autoencoders

Vineeth Reddy GUDA

 Overview of the Dataset

Context:
• Limited availability of financial datasets, especially in emerging domains like mobile money transactions.
• Synthetic dataset generated using the PaySim simulator to address this gap.
Content:
• PaySim simulates mobile money transactions based on real data from an African country.
• Dataset scaled down for Kaggle use, preserving the original transaction patterns.
• Fraudulent transactions are canceled, emphasizing the need for careful consideration during analysis.
Solution Overview

Utilizing Autoencoder for Anomaly Detection:

• Introduction of machine learning approach.
• Autoencoders are utilized for their ability to detect anomalies in
transaction data.
Understanding Autoencoders:
• Autoencoders consist of encoder and decoder layers.
• Encoder compresses input data, while decoder reconstructs it.
Detection via Reconstruction Error:
• Anomalies are identified based on reconstruction error.
• Higher error indicates potential fraudulent activity.
Learning to Distinguish Normal vs. Fraudulent Transactions:
• Autoencoder learns patterns to differentiate between normal and
fraudulent transactions.
• By training on diverse data, the model becomes adept at identifying
anomalies.
Libraries
numpy and pandas: Data manipulation and analysis.

keras: For building and training neural networks.

matplotlib and seaborn: Plotting and visualization.

pickle: Serialization of Python objects.

scikit-learn: Evaluation metrics for model

performance.
keras.backend: Access to TensorFlow backend
operations.
 Fraud Distribution:
Data Preprocessing • Majority (99.87%) of transactions are non-
fraudulent, with only 0.13% being fraudulent.
• Indicates significant class imbalance, posing a
challenge for model training.
Data Preprocessing:
• Log Transformation: Applied to numerical features
to handle skewness and improve distribution.
• Scaling: Normalized numerical features to a range
between 0 and 1 for uniformity.
• One-Hot Encoding: Converted categorical 'type'
variable to numerical format for model
compatibility.
Sample Preprocessed Data:
• Displayed sample rows of preprocessed data,
demonstrating transformed features and encoded
variables.
• Data is now ready for model training and evaluation,
enhancing interpretability and model performance.
 Autoencoder Model Architecture
Model Architecture:
• Input layer with 9 features representing transaction attributes.
• Encoder layer with 16 neurons, followed by a code layer with 8 neurons,
compressing input data into a lower-dimensional representation.
• Decoder layer with 16 neurons reconstructs the compressed representation.
• Output layer reconstructs the original 9 features using sigmoid activation.
Parameter Summary:
• Total trainable parameters: 593.
• Parameters include weights and biases of each layer in the autoencoder
architecture.
• No non-trainable parameters present.
Model Summary:
• Provides an overview of the layers, their output shapes, and the total number of
parameters.
• Highlights the simplicity and efficiency of the autoencoder architecture for fraud
detection tasks.
 Training the Autoencoder Model
Model Compilation:
• Loss function: Binary cross-entropy, suitable for binary classification tasks like fraud detection.
• Optimizer: Adam optimizer, known for its efficiency and adaptability.
• Metrics: Accuracy is tracked during training for evaluation purposes.
Callbacks:
• ModelCheckpoint: Saves the best performing model weights during training to "autoencoder_fraud.h5" file.
• TensorBoard: Logs training progress and writes logs for visualization in the './logs' directory.
Training Process:
• Dataset consists of 63,62,620 samples, with 20% used for validation (1,272,524 samples).
• Model trained for 10 epochs with a batch size of 128, shuffled before each epoch.
• Each epoch shows loss and accuracy metrics for both training and validation sets.
• Training time: Approximately 156 seconds per epoch.
 Model Evaluation:
• Loss Curves:
• Plots training and validation loss against epochs to visualize learning
progress.
• Interpretation:
• Training Loss: Decreases steadily, indicating improved reconstruction
of normal transactions.
• Validation Loss: Follows a similar trend, indicating good
generalization without overfitting.
• Convergence Analysis:
• Both curves converge towards the end of 10 epochs, indicating stable
model performance.
• Model Performance:
• Close alignment of curves demonstrates effective learning and
identifies fraudulent activities accurately.
• Overall decreasing trend signifies the effectiveness of autoencoder
architecture for fraud detection.
 Reconstruction Error Analysis
Reconstruction Error Calculation:
• Autoencoder predicts reconstructed transactions (train_x_predictions).
• Mean Squared Error (MSE) between original and reconstructed transactions
calculated.
• Error DataFrame created with MSE and true class labels.
Error Distribution:
• Mean reconstruction error: 8.35e-06, indicating overall low error in normal
transactions.
• Standard deviation: 6.43e-05, showing variability in reconstruction errors.
• Minimum error: 1.14e-11, representing accurately reconstructed transactions.
• Maximum error: 7.68e-02, indicating significant deviation from normal
behavior.
Interpretation:
• Majority of transactions have low reconstruction error, reflecting accurate
reconstruction of normal transactions.
• Higher errors may indicate anomalies or potentially fraudulent activities.
 Precision-Recall Analysis
Precision-Recall Curve:
• Visualizes the trade-off between precision and recall for various threshold values.
• Precision: Fraction of true positive predictions among all positive predictions.
• Recall: Fraction of true positive predictions among all actual positive instances.
• Demonstrates how precision and recall change with threshold adjustments.
Interpretation:
• Lowering the threshold increases recall (detecting more fraudulent transactions) but
decreases precision (increasing false positives).
• Precision-Recall curve aids in selecting an optimal threshold for practical
deployment of the fraud detection model.
Precision and Recall vs. Threshold:
• Plots precision and recall against different threshold values.
• Highlights the trade-off between precision and recall as the threshold varies.
• Intersection of curves assists in selecting a suitable threshold balancing the trade-off
between metrics.
 Confusion Matrix Analysis

Confusion Matrix:

• Provides visual representation of model predictions

compared to actual class labels.

Interpretation:

• Majority of normal transactions correctly identified

(6,354,406 true negative).
• Presence of false negatives (8,213) indicates potential
missed fraudulent transactions.
• Single false positive observed (normal transaction
misclassified as fraudulent).
Confusion Matrix
Analysis
Conclusion

Key Findings: Implications: Future Directions:

ML techniques, particularly the ML-driven fraud detection offers Continued refinement and
autoencoder model, show cost savings and enhances exploration of ML algorithms.
promise in fraud detection. customer trust. Application of findings to real-
Achieved commendable accuracy, Potential for further optimization world fraud detection systems.
precision, and recall rates. and exploration of advanced
algorithms.
Thank you…

Vineeth Reddy GUDA