Fraud Detection with Machine Learning

This project implements a machine learning pipeline to detect fraudulent credit card transactions. It includes dimensionality reduction via clustering, classification using a weak learner (decision stump), and iterative refinement with AdaBoost.

Overview

Given a set of transaction summaries and known labels (clean or fraud), this system builds a model to predict whether new transactions are fraudulent. Each transaction records spending at a set of physical locations (shops, restaurants, etc.). The pipeline consists of:

1. Clustering: Groups nearby map locations using Euclidean distance to reduce the dimensionality of transaction data.
2. Weak Learning: Trains a simple classifier (decision stump) to make predictions slightly better than random guessing.
3. Boosting: Iteratively refines predictions using AdaBoost, focusing on inputs that previous learners misclassified.

File Descriptions

Clustering.java

• Implements k-means-style clustering.
• Input: Point2D[] (map locations), integer k (number of clusters).
• Output: A mapping from m-dimensional transaction summaries to k-dimensional cluster summaries.

WeakLearner.java

• Implements a decision stump classifier.
• Trained on reduced transaction data (n x k), labels (0 or 1), and sample weights.
• Outputs a binary predictor to classify new samples as clean or fraud.

BoostingAlgorithm.java

• Implements AdaBoost using decision stumps.
• Input: original transaction data (n x m), Point2D[] locations, integer k (number of clusters).
• Internally:
  1. Reduces dimensionality using Clustering.
  2. Trains weak learners iteratively.
  3. Adjusts sample weights based on misclassification.
  4. Aggregates predictions using majority vote.

Data Format

• Transaction Summary: int[] of length m, representing spending at each location.
• Label: 0 (clean) or 1 (fraud).
• Map Locations: Point2D[] coordinate array.
• Weights: double[] of length n; initialized to 1/n and updated each round.

Prediction Process

To classify a new transaction:

1. Reduce its dimension using the Clustering object.
2. Get predictions from all trained weak learners.
3. Use majority voting to decide the final label.
   • In case of a tie, predict 0 (clean).

Dependencies

• Java 8 or later
• algs4.jar from the Princeton Algorithms library

  • Required for the Point2D data type and utility classes.

  • Add it to your project’s classpath during compilation and execution:

    javac -cp .:algs4.jar *.java
    java -cp .:algs4.jar BoostingAlgorithm