Feature-Optimized Quantum Classification Model for Financial Transaction Anomalies

Technical Summary: Feature-Optimized Quantum Classification Model for Financial Transaction Anomalies

Problem Statement
The exponential growth of digital financial transactions has created an urgent demand for intelligent fraud detection systems capable of identifying subtle, non-linear patterns that traditional machine learning models often miss. Conventional approaches frequently struggle with the high-dimensional, complex correlations inherent in financial data, particularly when distinguishing between legitimate and fraudulent activities. This study addresses the challenge of improving fraud detection accuracy by integrating classical data preprocessing with Quantum Machine Learning (QML) techniques to enhance feature representation and classification performance.

Methodology
The proposed framework follows a hybrid classical-quantum pipeline, beginning with rigorous data preprocessing and culminating in a Quantum Support Vector Machine (QSVM) classification.

1. Data Preprocessing and Feature Engineering:

   • Dataset: The study utilizes the "Synthetic Financial Datasets for Fraud Detection (Cleaned)" from Hugging Face, containing approximately 6.36 million synthetic transaction records with 11 attributes. The dataset exhibits a significant class imbalance, with fraudulent transactions comprising only ~0.13% of the total.
   • Handling Missing Values: The dataset is explicitly noted to contain zero missing values. While the methodology describes Mean, Median, and Mode Imputation strategies as part of the general framework, the specific dataset used in this study required no imputation as it was already clean.
   • Scaling and Normalization: To ensure uniform feature distribution and improve model convergence, the data undergoes Z-Score Standardization and Min-Max Scaling.
   • Feature Selection: Key numerical attributes (e.g., transaction amount, account balances before and after) are selected and scaled to the range $[0, \pi]$ to meet quantum rotation requirements.

2. Quantum Feature Mapping:

   • The preprocessed classical data is transformed into quantum states using the ZZFeatureMap. This feature map encodes data into a high-dimensional Hilbert space, utilizing entangling gates to capture pairwise feature interactions and non-linear correlations that are difficult for classical models to represent.
   • The mapping process involves defining the feature map dimensions, setting repetitions (reps=2), and utilizing linear entanglement.

3. Quantum Kernel Evaluation:

   • A Quantum Kernel is constructed using the ZZFeatureMap and a statevector simulator backend. This kernel evaluates the pairwise similarity of quantum-encoded samples, generating a kernel matrix that quantifies the correlation strength between data points in the transformed quantum space.

4. Classification (QSVM):

   • A Quantum Support Vector Machine (QSVM) is trained using the precomputed quantum kernel matrix. The dataset is split into 70% training and 30% testing (with stratification).
   • The classifier is trained on the training kernel matrix ($K_{train}$) and evaluated on the test set using the cross-kernel matrix ($K_{test}$).

Key Contributions

• Hybrid Framework Integration: The paper presents a complete pipeline that seamlessly integrates classical data engineering (scaling) with quantum feature mapping and kernel-based classification.
• Implementation of ZZFeatureMap: The study demonstrates the specific application of the ZZFeatureMap for financial fraud detection, showing how it transforms transaction data into a quantum Hilbert space to improve class separability.
• Benchmarking on Synthetic Data: The research provides a structured evaluation of a QSVM approach on a large-scale, realistic synthetic financial dataset, offering a baseline for future QML applications in finance.

Results
The proposed Quantum-Enhanced Fraud Detection Framework achieved the following performance metrics on the test set:

• Accuracy: 91%
• Precision (Fraud): 0.88
• Recall (Fraud): 0.90
• F1-Score (Fraud): 0.89
• Precision (Normal): 0.92
• Recall (Normal): 0.91
• F1-Score (Normal): 0.91

The results indicate a balanced performance with minimal false positives and false negatives, suggesting the model generalizes well despite the class imbalance.

Significance and Claims
The author claims that this work represents an initial phase in integrating quantum-enhanced computation into financial fraud detection. The study validates the hypothesis that Quantum Feature Mapping, when combined with Quantum Kernel-based SVM, can outperform traditional classifiers on structured financial data.

The significance lies in the demonstrated ability of the quantum approach to capture complex, entangled relationships between features, leading to superior discrimination between fraudulent and legitimate transactions. The paper posits that this hybrid quantum-classical approach offers a promising direction for future research, paving the way for scalable, secure, and intelligent fraud prevention systems that leverage the generalization and robustness of Quantum Machine Learning. The author emphasizes that the choice of feature map and hyperparameters is crucial, aligning with recent literature on the practical advantages of quantum kernels over classical baselines.