PaperlayerView exact revision on arXiv

Source-linked AI summary

Mixed Quantum-Classical Method For Fraud Detection with Quantum Feature Selection

Michele Grossi, Noelle Ibrahim, Voica Radescu, Robert Loredo, Kirsten Voigt, Constantin Von Altrock, Andreas Rudnik

arXiv:2208.07963v1quant-phcs.LG

TL;DR

Payment fraud detection must balance fraud losses with false positives, while fraud schemes are becoming more complex. This paper compares quantum and classical methods on real card-payment data, explores quantum feature selection, and tests a mixed ensemble. The authors report improved accuracy for the mixed quantum-classical method using QSVM’s complementary feature-space exploration on a drastically reduced dataset.

Abstract

from arXiv · show

This paper presents a first end-to-end application of a Quantum Support Vector Machine (QSVM) algorithm for a classification problem in the financial payment industry using the IBM Safer Payments and IBM Quantum Computers via the Qiskit software stack. Based on real card payment data, a thorough comparison is performed to assess the complementary impact brought in by the current state-of-the-art Quantum Machine Learning algorithms with respect to the Classical Approach. A new method to search for best features is explored using the Quantum Support Vector Machine's feature map characteristics. The results are compared using fraud specific key performance indicators: Accuracy, Recall, and False Positive Rate, extracted from analyses based on human expertise (rule decisions), classical machine learning algorithms (Random Forest, XGBoost) and quantum based machine learning algorithms using QSVM. In addition, a hybrid classical-quantum approach is explored by using an ensemble model that combines classical and quantum algorithms to better improve the fraud prevention decision. We found, as expected, that the results highly depend on feature selections and algorithms that are used to select them. The QSVM provides a complementary exploration of the feature space which led to an improved accuracy of the mixed quantum-classical method for fraud detection, on a drastically reduced data set to fit current state of Quantum Hardware.

I. INTRODUCTION

Payment fraud prevention must balance fraud losses against false positives, while fraud schemes are becoming more complex. The paper compares quantum and classical approaches on real card-payment data and investigates quantum feature selection and mixed ensembles.

I. INTRODUCTION

  • Payment fraud prevention balances monetary losses from fraud against false positives that can harm client experience, while criminals are using increasingly complex schemes.The paper notes that some criminals use AI/ML to augment their attacks.Source: §I · p. 1
  • The study compares quantum feature selection with classical selection, evaluates QSVM against Random Forest and XGBoost on real card-payment data, and tests mixed quantum/classical ensembles.The stated aim is to assess their impact on classifier performance.Source: §I · p. 2

A. METHODOLOGY

  • The study uses IBM Safer Payments as its experimentation platform and exports training data from it for classical and quantum models.The authors describe external model import as making integration feasible, while noting that latency considerations are outside the paper’s scope.Source: A. METHODOLOGY · p. 2

B. PAYMENT FRAUD PREVENTION KPIS

The section describes fraud-prevention metrics and the real-world payment data context. It distinguishes monetary fraud-prevention measures from classifier evaluation metrics and notes a threshold-based transaction decision.

B. PAYMENT FRAUD PREVENTION KPIS

  • Fraud prevention is evaluated using hit rate—the share of confirmed fraud value declined—and false alarm ratio, the ratio of false alerts to true alerts.False alerts can disrupt customer payments and trigger manual interaction, largely independently of transaction amount.Source: B. PAYMENT FRAUD PREVENTION KPIS · p. 2
  • Accuracy measures correct predictions but is meaningful only on balanced data; the paper uses a modified ROC curve whose x axis is false alarm ratio.The text describes AUC as a measure of separability.Source: B. PAYMENT FRAUD PREVENTION KPIS · p. 2
  • A classifier score is converted into a real-time decline-or-approve decision by applying a threshold.Scores may be ordinal or represent predicted fraud probability.Source: B. PAYMENT FRAUD PREVENTION KPIS · p. 2

II. INPUT DATA SET

  • The dataset contains real-world payment transactions from a European cross-border processing portfolio, with cardholder demographic and transaction information available for enrichment.The passage reports about 80% debit and 20% credit transactions, 2.4 million total transactions, and about 3,000 marked frauds.Source: §II · p. 2, A. TRANSACTION AND CUSTOMER DATA · pp. 2–3

B. ENGINEERED FEATURES

The paper describes engineered behavioral features and separate data preparations for rule-based, classical machine-learning, and quantum use cases. The quantum setting requires substantial dimensionality reduction because of hardware limits.

B. ENGINEERED FEATURES

  • Behavioral profiles aggregate transaction counts and frequencies over time windows, and the authors describe them as strongly discriminating between fraudulent and non-fraudulent transactions.The initial 48 attributes include mixed data types, requiring treatment of categorical data for machine-learning classifiers.Source: B. ENGINEERED FEATURES · p. 3

C. DATASETS FOR USE CASES

  • The paper frames dimensionality reduction as necessary for QML because quantum hardware is limited in qubits and error rates.The stated aim is to reduce dimensions while maximally preserving model accuracy.Source: C. DATASETS FOR USE CASES · p. 3
  • The study uses three data references: full-dimensional data for expert-assisted rules, balanced data for classical ML, and drastically reduced, normalized data for quantum and direct classical–quantum comparison.The reduced quantum dataset uses multiple trials to address bias from heavy under-sampling.Source: C. DATASETS FOR USE CASES · p. 3

III. CLASSICAL FRAUD DETECTION MODEL

  • Fraud-prevention decision models may be created with conventional rules or machine-learning techniques.Source: §III · p. 3

A. HUMAN AND INTERACTIVE EXPERT METHOD

The expert Rule Generator trades fraud capture against false alarms, while classical classifiers are evaluated against it on complete and under-sampled data. Classical classifiers rank features differently, and the under-sampled comparison is reported as comparable to the full-data results.

A. HUMAN AND INTERACTIVE EXPERT METHOD

  • The Rule Generator’s Hit Rate can reach 65% with 25 non-fraudulent payments intercepted per fraud, or 30% with five; its training curve sits above its test curve, indicating overfitting.The choice between these KPI trade-offs is left to the processor or bank.Source: A. HUMAN AND INTERACTIVE EXPERT METHOD · p. 3

B. CLASSICAL MACHINE LEARNING CLASSIFIERS

  • Classical preprocessing removed highly correlated features, expanded categorical attributes from 48 to 69 features, split training and test data, and under-sampled genuine transactions across five random trials.The trials preserve the same fraudulent-to-genuine transaction ratio.Source: 1) Data Preparation for Classical… · p. 4
  • XGBoost and Random Forest assign different feature-importance orders, motivating a different view of the feature space through alternative selection methods.The authors connect this observation to studying quantum feature selection as a complement to classical methodology.Source: 2) Classical Machine Learning with… · p. 4
  • At a 50% Hit Rate, the complete-data comparison reports a False Alarm Rate above 1:50; XGBoost appears to outperform Random Forest, though their Hit Rate uses record counts versus the Rule Generator’s monetary basis.Figure 2 compares Hit Rate and False Alarm Ratio for XGBoost, Random Forest, and Rule Generator.Source: 2) Classical Machine Learning with… · p. 5, Figure 2 · p. 5
  • Under-sampled XGBoost results were comparable to the earlier full-data results after scaling false alarms by 500, supporting the under-sampling procedure for reduced-data experiments.The reported Hit Rate for the Rule Generator is monetary, while classical-model Hit Rate is based on record count.Source: 3) Classical Machine Learning with… · p. 5, Figure 3 · p. 5

IV. QUANTUM MACHINE LEARNING FOR FRAUD DETECTION

The section frames QSVM kernels as a route to better fraud classification and identifies feature reduction as a practical challenge. It also notes that FAMD components overlap and names Qiskit as the software used.

IV. QUANTUM MACHINE LEARNING FOR FRAUD DETECTION

  • The goal is to find a quantum kernel that improves classification accuracy, but no general recipe for building such kernels is yet available.The section notes kernel alignment as an optimization technique for specific classes of quantum kernels.Source: §IV · p. 5
  • FAMD shows total overlap between components, limiting its use for reducing the feature set.The figure depicts relations between FAMD components.Source: Figure 4 · p. 6
  • The work uses QSVM in two steps: select features to reduce data dimensionality, then derive fraud KPIs from a quantum machine-learning model.The experiments use the Qiskit quantum software package.Source: §IV · pp. 5–6

A. QUANTUM FEATURE IMPORTANCE SELECTION

The feature-selection method searches QSVM feature combinations directly because classical feature rankings may not preserve relationships a quantum model could detect. It grows feature sets iteratively, comparing quantum feature maps by accuracy and AUC.

A. QUANTUM FEATURE IMPORTANCE SELECTION

  • Under-sampling is used to reduce records for quantum hardware, and inputs are normalized to [−1, 1] for mapping to quantum rotations.The authors tested the method on ideal and noisy simulators and then on real hardware with error mitigation.Source: A. QUANTUM FEATURE IMPORTANCE… · p. 6
  • PCA was unsuitable for the mostly binary or categorical data, while FAMD’s first two components overlapped completely and showed no discrimination.The authors also evaluated feature importance from XGBoost and Random Forest on the full 2.4-million-record dataset.Source: A. QUANTUM FEATURE IMPORTANCE… · p. 6
  • Classical feature-importance rankings may bias QSVM outcomes, motivating a quantum search using feature maps and kernels to select features.The authors report that different classifiers prefer different features.Source: A. QUANTUM FEATURE IMPORTANCE… · p. 6
  • The search iteratively increases feature count from three of 69, following a feed-forward feature-selection approach based on accuracy or AUC.Feature selection is integrated into the paper’s fraud-detection framework.Source: A. QUANTUM FEATURE IMPORTANCE… · p. 6, Figure 5 · p. 7
  • The procedure selects the best three of 69 features, then adds features through further permutations until improvement saturates or a qubit–accuracy trade-off is reached.Accuracy is the selection criterion, with AUC also stored; the search begins with thousands of combinations.Source: A. QUANTUM FEATURE IMPORTANCE… · p. 7
  • The search tests Z and ZZ feature maps at depths 1 and 2, selecting the best feature combination at each stage by accuracy.Figure 6 shows the spread of accuracy values across combinations for sets of three to six features.Source: A. QUANTUM FEATURE IMPORTANCE… · p. 7, Figure 6 · p. 7

B. QSVM RESULTS

QSVM feature selection improves outcomes over QSVM using features selected by classical models. Accuracy also rises with additional features, and the authors report preference for more-entangled maps, particularly ZZ depth 2.

B. QSVM RESULTS

  • Accuracy improves as features are added, and the authors report a preference for the entangled ZZ feature map at depth 2.They state that more-entangled maps perform better but are harder to simulate on classical hardware.Source: B. QSVM RESULTS · p. 8
  • QSVM selects features different from classical algorithms, providing a complementary scan of the feature space.Table 5 lists the seven best QSVM-selected features across ZZ-map and depth choices; its defining KPI is accuracy, with best KPIs for ZZ depth 2.Source: B. QSVM RESULTS · p. 8, Table 5 · p. 8
  • The selected features show low overlap, though identifying independent meaningful features is difficult because many were engineered from the same raw inputs.The correlation matrix is presented in Table 8.Source: B. QSVM RESULTS · p. 8
  • QSVM feature selection improved outcomes over QSVM with XGBoost- and Random Forest-selected features on the same datasets (1,500 training, 1,000 testing records); Table 6 reports average accuracy and AUC on balanced data.Source: B. QSVM RESULTS · p. 8, Table 6 · p. 8
  • The seven-feature circuit has depth around 70, and reducing circuit depth for hardware use is future work.Figure 10 illustrates a circuit for three features.Source: B. QSVM RESULTS · p. 9, Figure 10 · p. 9

V. MIXED QUANTUM-CLASSICAL METHOD FOR FRAUD DETECTION

The paper explores QSVM’s complementarity with classical machine learning. Its hybrid approach uses classical data and function learning with a quantum classification algorithm.

V. MIXED QUANTUM-CLASSICAL METHOD FOR FRAUD DETECTION

  • The section focuses on complementarity between QSVM and classical machine learning.The paper describes this as a complementarity exploration, rather than asserting quantum methods replace classical ones.Source: §V · p. 9
  • The approach uses QSVM to explore quantum-classical complementarity in fraud detection.The section does not report a quantitative result for the mixed method.Source: §V · p. 9
  • In the hybrid setup, data and function learning are classical, while the classification algorithm is quantum.The quantum computer computes the training-set kernel matrix once; a classical computer solves the dual optimization problem for the hyperplane and support vectors.Source: §V · p. 9

A. APPROACH

The approach uses disagreements between quantum and classical classifiers to train a metaclassifier that predicts which classifier to trust for a transaction.

A. APPROACH

  • Transactions on which the quantum and classical classifiers disagree in the balanced training set form an additional dataset for metaclassifier training.The classifiers are first trained on the training portion of the balanced dataset.Source: A. APPROACH · p. 9
  • The metaclassifier uses transaction features to decide which classifier to trust when their predictions differ.It may use any dataset features and can take different forms.Source: A. APPROACH · p. 9
  • A classical SVM serves as metaclassifier for fully optimized XGBoost, while logistic regression serves for XGBoost without optimization.Because few training points had classifier disagreements, simple metaclassifiers performed best.Source: A. APPROACH · p. 9

B. RESULTS AND COMPARISONS

The results describe complementary quantum and classical predictions and use a metaclassifier to improve performance; the supplied passages do not report comparative accuracy values.

B. RESULTS AND COMPARISONS

  • Quantum and classical classifiers disagree on 5.2% of training data and 5.5% of test data at a 0.5 threshold.The paper describes these differing predictions as complementary.Source: B. RESULTS AND COMPARISONS · pp. 9–10
  • The metaclassifier exploits complementary predictions to increase performance, with accuracy comparisons framed for separate and mixed models on the balanced dataset.Tables 8 and 9 are identified as comparisons of model accuracy for separate and mixed models, respectively; their captions do not state the values or a winner.Source: B. RESULTS AND COMPARISONS · p. 10, Table 9 · p. 10, Table 8 · p. 10
  • The study reports using the IBM Qiskit simulator to employ QSVM and measure accuracy and AUC across different use cases.Source: B. RESULTS AND COMPARISONS · p. 10

VI. SUMMARY

The summary presents quantum feature selection and a mixed quantum-classical ensemble as complementary approaches for fraud classification, while noting that the reported results use a simulator.

VI. SUMMARY

  • Quantum Feature Importance Selection uses QSVM to select the most relevant features as the number of selected features increases.Source: §VI · p. 10
  • Quantum feature maps with more entanglement systematically provide better KPIs in the reported experiments.Source: §VI · p. 10
  • The workflow requires careful preprocessing, including data-type considerations and under-sampling, before quantum processing.Source: §VI · p. 10
  • Quantum classifiers identify patterns described as difficult for classical algorithms to detect, and the ensemble can improve the final model's KPI while balancing false positives and false negatives.Source: §VI · p. 10
  • The reported results come from a simulated quantum computer; real-hardware implementation is left to a different manuscript.Source: §VI · p. 10
Loading 2208.07963v1…