Instance-dependent cost-sensitive learning for detecting transfer fraud

Sebastiaan Höppner, Bart Baesens, Wouter Verbeke, Tim VerdonckEur. J. Oper. Res 2022 · OpenReview

Abstract

This paper introduces an instance-dependent cost-sensitive learning framework designed specifically for transfer fraud detection. By defining an instance-dependent cost matrix and deriving an optimal threshold, we propose two novel algorithms: cslogit (Lasso-regularized cost-sensitive logistic regression) and csboost (cost-sensitive gradient tree boosting). These methods directly minimize the financial loss associated with fraud, rather than optimizing for standard statistical metrics. Experimental results on both public and proprietary datasets demonstrate that these models significantly increase financial savings compared to traditional classification methods.

Introduction

In the financial sector, the cost of misclassification in fraud detection is highly asymmetric and instance-dependent. Missing a $10,000 fraudulent transfer (a False Negative) is far more damaging than missing a $10 one. Conversely, a False Positive—flagging a legitimate transaction—incurs a fixed administrative cost for investigation. Despite this, most machine learning models are trained to optimize for metrics like AUC or $F_1$ scores, which treat every transaction as having equal weight. This paper bridges the gap between statistical accuracy and business utility by building models that "think" in terms of financial impact.

The Cost-Sensitive Framework

The foundation of this work is the Average Expected Cost (AEC) objective function, which incorporates the specific costs associated with different classification outcomes.

Instance-dependent cost matrix for transfer fraud detection.

As illustrated in the cost matrix above, the cost of a False Negative for transaction $i$ is its amount $A_i$. The cost of a False Positive is a fixed administrative fee $c_f$. Notably, even a True Positive (correctly identified fraud) incurs the cost $c_f$ due to the necessity of manual investigation. From this matrix, an optimal instance-dependent threshold is derived: a transaction should be flagged if its fraud probability $s_i$ exceeds $c_f / A_i$. This mathematical result dictates that high-value transactions require a lower probability threshold to justify an alert.

Proposed Methods

To optimize the AEC directly, the authors propose two methods:

  1. cslogit: This method adapts logistic regression to minimize the AEC. To handle the potentially unstable optimization surface of the cost-sensitive objective, a Lasso ($\ell_1$) penalty is added. This regularization stabilizes the gradient-based search and performs automatic feature selection.
  2. csboost: This method integrates the AEC into the Gradient Tree Boosting framework (specifically XGBoost). By deriving custom first-order (gradient) and second-order (Hessian) derivatives for the cost-sensitive loss, the algorithm builds decision trees that prioritize reducing financial loss.

Results and Discussion

The proposed models were evaluated on the Kaggle Credit Card dataset and a proprietary bank dataset. The primary performance metric was Savings, representing the percentage of total potential fraud loss prevented.

  • Performance: On the Kaggle dataset, cslogit achieved 66.46% expected savings, a substantial improvement over the 46.16% achieved by standard logistic regression.
  • Economic Impact: In the real-world bank dataset, the use of cslogit resulted in an average cost reduction of €0.62 per transaction. For a large-scale operation, this translates to tens of thousands of Euros in daily savings.
  • Metric Divergence: A key finding was that models optimized for cost often yielded lower $F_1$ scores than standard models. This highlights a critical "metric divergence": optimizing for statistical precision can lead to sub-optimal financial performance in production environments.

Conclusion

Instance-dependent cost-sensitive learning provides a more rational approach to fraud detection by aligning model objectives with business goals. The proposed cslogit and csboost algorithms offer practical, efficient solutions for minimizing financial loss. Future research will focus on incorporating variable costs for customer friction and extending these principles to deep learning architectures.

Reference: Höppner, S., Baesens, B., Verbeke, W., & Verdonck, T. (2020). Instance-Dependent Cost-Sensitive Learning for Detecting Transfer Fraud. Preprint submitted to European Journal of Operational Research.

Summary updated

Abstract

Highlights•An instance dependent cost matrix for transfer fraud detection is presented.•An instance-dependent threshold is derived.•Two instance-dependent cost-sensitive methods are proposed.•They create the detection model that minimizes the financial loss due to fraud.•Useful for any classification problem involving an instance-dependent cost-matrix.
BibTeX
@article{hoppner2022instance,
  author = {Sebastiaan Höppner and Bart Baesens and Wouter Verbeke and Tim Verdonck},
  title = {{Instance-dependent cost-sensitive learning for detecting transfer fraud}},
  journal = {Eur. J. Oper. Res.},
  year = {2022},
  url = {https://openreview.net/forum?id=SRlIkv0j5z},
}

Loading related content...

Figure