Financial Fraud Analysis with KNIME

A comprehensive data mining and machine learning project built in KNIME Analytics Platform aimed at identifying and analyzing fraudulent transactions within the synthetic PaySim mobile money dataset.

📖 Project Overview

This project implements an end-to-end data analysis workflow in KNIME to detect suspicious and fraudulent financial activities. The workflow encompasses data loading, exploratory data analysis (EDA), data preprocessing, feature engineering, and training multiple machine learning models. The goal is to evaluate the effectiveness of different models within the KNIME ecosystem for the task of financial anomaly detection.

---

📊 Dataset

Name: Synthetic Financial Datasets For Fraud Detection
Source: Available on Kaggle.
Description: A realistic synthetic dataset simulating mobile money transactions, designed for fraud detection research.
Size: 6,362,620 transactions
Key Features: step, type, amount, nameOrig, oldbalanceOrg, newbalanceOrig, nameDest, oldbalanceDest, newbalanceDest, isFraud, isFlaggedFraud.

🧩 KNIME Workflow Overview

The main workflow (project.knwf) is structured into several key sections:

1. Data Input & Exploration: Using a File Reader node to load the data, followed by Rule Engine, Statistics, and Bar Chart nodes to understand data distribution, missing values, and the concentration of fraud.
2. Data Preprocessing: Filtering data (using a Row Filter) to focus only on TRANSFER and CASH_OUT transactions. Creating a new feature hour_of_day from the step column using a Math Formula node.
3. Model Training & Validation: The workflow splits into parallel branches for different models:
   • K-Means Clustering: Using the k-Means node for unsupervised learning.
   • Supervised Models: Using the Partitioning node to split data and various learner/predictor nodes:
     • Decision Tree
     • Naïve Bayes
     • Random Forest
   • Model evaluation is performed using Cross-Validation meta-nodes and scored using Scorer and Numeric Scorer nodes to calculate accuracy, precision, recall, and F1-score.
4. Results Visualization: Using ROC Curve, Scatter Plot, and View nodes to visualize model performance and analysis results.

🎯 Key Insights from EDA

• Fraud is exclusively present in TRANSFER and CASH_OUT transaction types.
• Extreme class imbalance: Only 8,213 fraudulent transactions out of 6+ million.
• No direct correlation was found between amount, time of day (step), or specific users and fraudulent activity.

---

📈 Results & Performance

The performance of all implemented models is summarized in the table below. The key metrics are Error Count and Accuracy, with a detailed breakdown of prediction types (True/False Positives/Negatives).

Model	Error Count	True Positive	True Negative	False Positive	False Negative	Accuracy
k-Means Clustering	1,364,345	5,469	1,400,595	1,361,601	2,744	50.75%
Naïve Bayes	1,642	0	2,762,196	0	8,213	99.70%
Decision Tree (Outliers Removed)	865	232	2,521,287	168	4,160	99.82%
Decision Tree (Outliers Not Removed)	1,345	1,899	2,761,784	412	6,314	99.75%
Random Forest (Outliers Removed)	876	10	2,521,454	1	4,382	99.82%
Random Forest (Outliers Not Removed)	1,411	1,226	2,762,125	71	6,987	99.74%

---

⚠️ Challenges & Limitations

1. Class Imbalance: The extreme rarity of fraud cases leads to models biased towards the majority class.
2. Pattern Repetition: Artificial patterns in the data can cause overfitting.
3. Limited Feature Set: Lack of user historical data limits behavioral analysis.

---

🧑‍🎓 Author

Fatemeh Mousavi
Student at IUT
Course: Data Mining
Semester: Spring 2024-2025