Fraud Detection Work

Final-year project: Machine Learning for Credit Card Fraud Detection by Michael Brozhko.
End-to-end pipeline covering dataset assembly, preprocessing, feature engineering, dimensionality checks, model training with hyperparameter search, and evaluation on original vs oversampled data.

---

Table of Contents

• Motivation
• Dataset
• Setup
• Workflow
• Feature Engineering
• Dimensionality & Selection
• Models & Tuning
• Results (Summary)
• How to Run
• Key Functions
• Notes & Assumptions

---

Motivation

Fraud datasets are highly imbalanced and may come pre-split in ways that bias results.
This project recombines the original train/test CSVs and performs a fresh randomized split to ensure fair evaluation and reproducibility.

---

Dataset

Two CSVs are combined into a single combined_csv.csv, then split with train_test_split.
Columns include transaction metadata (time, amount, merchant/customer geolocation), demographics, and the target label is_fraud.

Path note: Update the dataset folder path before running the combine step.

---

Setup

Python 3.10+ recommended

pip install -r requirements.txt

Core libraries: pandas, numpy, scipy, matplotlib, scikit-learn, imbalanced-learn, xgboost.

Workflow

• Combine CSVs → combined_csv.csv
• Load & inspect datatypes and nulls (none initially)
• Preprocess
• Drop high-cardinality / ID-like or free-text fields (e.g., cc_num, trans_num, names, streets)
• Convert/expand datetime → hour, day (0–6), month
• Derive age from dob
• Feature engineering (see below)
• Encode categorical features (category, gender) via One-Hot Encoding
• Train/test split (random_state=42)
• Outlier guardrails (IQR checks for age, amt; cap age at 98 in this study)
• PCA diagnostics to understand variance contribution
• Modeling with hyperparameter search (GridSearchCV, recall-focused scoring)
• Evaluation on:
  • Original (unsampled) data
  • Random over-sampled data (10:1 majority:minority)

Feature Engineering

• Temporal: hour, day, month
• Age: age = current_year − dob.year
• Geo distance: merchant_customer_distance = sqrt((lat − merch_lat)^2 + (long − merch_long)^2)
• Final numeric subset used for core models: ['amt','zip','lat','long','age','hour','day','month','merchant_customer_distance']

Dimensionality & Selection

PCA scree/variance analysis guided a compact numeric feature set (above).

One-hot features for category/gender were explored; final primary models emphasize the numeric subset for robustness and speed.

Models & Tuning

All models run in imblearn pipelines with scaling and are tuned via GridSearchCV, optimizing recall of the fraud class (1).

KNN: n_neighbors ∈ {4,5,6}, weights='distance'

XGBoost: depth, learning rate, estimators, L1/L2 regularization

Random Forest: estimators, max depth

MLP: hidden sizes, activation (tanh/relu), max_iter

Scoring uses make_scorer(recall_score, pos_label=1) to prioritize catching fraud.

Results (Summary)

Unsampled (original imbalance)

XGBoost — Accuracy: 0.9990, Fraud Recall: 0.85, Precision: 0.96

Random Forest — Accuracy: 0.9987, Fraud Recall: 0.80, Precision: 0.95

KNN — Accuracy: 0.9985, Fraud Recall: 0.82, Precision: 0.89

MLP — Accuracy: 0.9966, Fraud Recall: 0.49, Precision: 0.79

RandomOverSampler (10:1)

XGBoost — Accuracy: 0.9988, Fraud Recall: 0.87, Precision: 0.91

Random Forest — Accuracy: 0.9989, Fraud Recall: 0.87, Precision: 0.92

KNN — Accuracy: 0.9981, Fraud Recall: 0.84, Precision: 0.81

MLP — Accuracy: 0.9944, Fraud Recall: 0.85, Precision: 0.48

Takeaways

Resampling improves minority recall (fraud detection) with negligible impact on overall accuracy.

Tree-based models (XGBoost / RF) provide the best recall–precision balance.

KNN is competitive but heavier at scale; MLP needs further tuning/regularization for precision.

How to Run

1) Combine CSVs (adjust path)

import os, glob, pandas as pd os.chdir("/path/to/Dataset") combined_csv = pd.concat([pd.read_csv(f) for f in glob.glob("*.csv")]) combined_csv.to_csv("combined_csv.csv", index=False)

2) Load, preprocess, encode, split (see script/notebook)

import pandas as pd df = pd.read_csv("combined_csv.csv")

... feature engineering, OHE, split → X_train, y_train, X_test, y_test

3) Train/evaluate

from imblearn.over_sampling import RandomOverSampler runPipelines(X_train, y_train, X_test, y_test, sampler=None) # unsampled runPipelines(X_train, y_train, X_test, y_test, sampler=RandomOverSampler(random_state=42, sampling_strategy=0.1))

Key Functions

pca_analysis(features) — standardizes, fits PCA, plots scree, prints explained variance ratio.

modelPerformance(pipe, name, parameters, X_train, y_train, X_test, y_test) — grid search (recall scorer), fits best model, prints train/test metrics, plots precision/recall/F1.

runPipelines(X_train, y_train, X_test, y_test, sampler) — optional re-sampling, then runs/tunes KNN, XGBoost, RF, and MLP.

Notes & Assumptions

Reproducibility: fixed random_state=42 where applicable.

Leakage avoidance: dropped identifiers and free-text personal fields.

Scaling: standardization inside pipelines (keeps setup consistent across models).

Imbalance handling: RandomOverSampler at 0.1 (≈10:1) gave the best recall/precision trade-off in this study.

Plots: classification bar charts generated with matplotlib.