Fraud Detection in Bank Transactions [Our Reference Paper.]

This repository contains the solution for the Innotech AI Competition, focused on detecting fraudulent activities in bank transactions using advanced machine learning techniques. The project applies state-of-the-art algorithms to identify potentially fraudulent transactions with high accuracy.

Table of Contents

• Dataset
• Methodology
• Model Architecture
• Results
• Requirments
• Usage
• Refrences

Methodology

The approach to fraud detection in this project involves several key steps:

1. Data Preprocessing:

   • Handling missing values.
   • Feature engineering to create meaningful features from raw data.
   • Data normalization and scaling.

2. Exploratory Data Analysis (EDA):

   • Visualizing the distribution of features.
   • Identifying correlations between features.
   • Detecting any anomalies in the data.

3. Model Selection:

   • Comparison of different machine learning models including Random Forest, Gradient Boosting, and Neural Networks.
   • Use of cross-validation techniques to ensure model robustness.
   • Hyperparameter tuning to optimize model performance.

4. Evaluation Metrics:

   • Precision, recall, F1-score, and AUC-ROC curve were used to evaluate the model performance.
   • Special emphasis on minimizing false positives to avoid disrupting legitimate transactions.

Model Architecture

• Convolutions:
  • conv1: Edge Convolution (EdgeConv) for initial node feature aggregation.
  • conv2 & conv3: Transformer-based convolutions (TransformerConv) with skip connections for advanced feature extraction.
• Batch Normalization:
  • Applied after each convolutional layer to stabilize and accelerate training.
• Skip Connections:
  • Skip connections are integrated into the convolutional layers (conv2 and conv3) to retain information from previous layers.
• Linear Layers:
  • lin1: Reduces the dimensionality of the features before the final classification.
  • lin: Final classifier layer to output node class probabilities.
• Global Mean Pooling:
  • Aggregates node features into a graph-level representation.

Forward Pass

1. Node Embeddings:
   • Nodes are passed through a series of convolutional layers with skip connections to refine their embeddings.
2. Readout Layer:
   • Node features are aggregated at the graph level using global mean pooling.
3. Classification:
   • The pooled features are passed through linear layers to predict the class of each node.

Training

• Optimizer: Adam optimizer with a learning rate of 0.01.
• Loss Function: Negative Log-Likelihood Loss (NLLLoss) for classification.

Dataset

The dataset used for this project was provided by the Innotech AI Competition and contains a large number of bank transaction records, including both legitimate and fraudulent transactions. The data includes features such as transaction amount, transaction time, account age, and more.

• Dataset Source: Provided by Innotech AI Competition.
• Number of Transactions: 100K.
• Number of Features: 10.

Results

The model was evaluated using the test set provided by the competition, and it achieved the following results:

• Accuracy: 0.85
• Recall: 0.83
• F1-Score: 0.84

Usage

You can check the jupyter files out and find the solution. check them base on this order: 1-Data Loader.ipynb 2-model-with-node_attr.ipynb

Requirments

Pytorch
Pytorch geometric
Numpy
Pandas
Matplotlib

Refrences

• Yun, Seongjun, et al. "Graph Transformer Networks." International Conference on Learning Representations, 2020. Link to paper.