Paper: arXiv 2508.02283

Authors: Francis Boabang, Samuel Asante Gyamerah

Abstract

In insurance fraud prediction, handling class imbalance remains a critical challenge. This paper presents a novel multistage focal loss function designed to enhance the performance of machine learning models in such imbalanced settings by helping to escape local minima and converge to a good solution. Building upon the foundation of the standard focal loss, our proposed approach introduces a dynamic, multi-stage convex and nonconvex mechanism that progressively adjusts the focus on hard-to-classify samples across training epochs. This strategic refinement facilitates more stable learning and improved discrimination between fraudulent and legitimate cases. Through extensive experimentation on a real-world insurance dataset, our method achieved better performance than the traditional focal loss, as measured by accuracy, precision, F1-score, recall and Area Under the Curve (AUC) metrics on the auto insurance dataset. These results demonstrate the efficacy of the multistage focal loss in boosting model robustness and predictive accuracy in highly skewed classification tasks, offering significant implications for fraud detection systems in the insurance industry. An explainable model is included to interpret the results.

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper focuses on a modified loss function (multistage focal loss) which involves moderate mathematical formulation but primarily adapts existing concepts, while the empirical section details extensive experimentation on a real-world insurance dataset with standard ML metrics (accuracy, F1, AUC) and mentions preprocessing like SMOTE/ADASYN.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Improve Fraud Detection<br>in Class Imbalanced Data"] --> B{"Data Input"}
  B --> C["Real-World Auto Insurance Dataset<br>Skewed: Legitimate vs Fraud"]
  C --> D["Key Methodology:<br>Multistage Focal Loss Function"]
  D --> E["Computational Process:<br>Training ML Models with<br>Dynamic Stage Adjustments"]
  E --> F["Performance Metrics:<br>F1, Precision, Recall, AUC"]
  F --> G["Key Outcome:<br>Outperforms Standard Focal Loss<br>Explainable AI Included"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a structured three-stage training framework that uses a convex surrogate loss, an intermediate non‑convex loss, and the standard focal loss to improve fraud detection under severe class imbalance.

Source sentence
This paper proposes a structured three-stage training framework that integrates a convex surrogate of focal loss for stable initialization, a controlled non-convex intermediate loss to improve feature discrimination, and the standard focal loss to refine minority-class sensitivity.

Assumes 3

  • The convex surrogate loss remains convex when the weight wt is fixed and the focusing parameter γ is chosen so that 1 − γpt > 0. model
    Source sentence
    The convexity result therefore holds under the assumption that wt is fixed and that we analyse the loss as a function of pt
  • The dataset is highly imbalanced with far fewer fraudulent cases than legitimate ones. data
    Source sentence
    The dataset is usually highly skewed, with legitimate (non‑fraudulent) cases becoming more than the fraudulent ones. Formally, |{yi = 1}| ≪ |{yi = 0}|, resulting in bias…
  • Deep sequential (LSTM) models are used and their overall problem is not convex, but a convex surrogate in logit space provides a well‑conditioned initial phase. model
    Source sentence
    Thus, while the convex warm start confines the optimisation trajectory to a locally well‑conditioned region in logit space, the LSTM mapping remains intrinsically non‑convex.

Data

  • Proprietary auto‑insurance claims dataset · large U.S. insurer other proprietary
    Source sentence
    The empirical analysis uses a dataset of automobile insurance claims from a large U.S. insurer. The sample contains 39,981 records, each associated with a single customer, and 39 explanatory variables describing policy, vehicle, customer, and claim characteristics.

Costs

Not applicable The paper reports no trading or execution result.

Finds 2

  • The three‑stage schedule delivers the most balanced and reliable performance across accuracy, precision, recall, F1 score, and AUC, outperforming both convex‑only and non‑convex‑only baselines. performance
    Source sentence
    Using accuracy, precision, recall, F1 score, and AUC as evaluation metrics, we found that the three‑stage schedule delivers the most balanced and reliable performance
  • The staged transition from convex to non‑convex learning produces a sparse, well‑regularized SHAP attribution profile with moderate, bounded feature importance magnitudes. predictive-accuracy
    Source sentence
    The staged transition from convex to non‑convex learning led to selective and well‑regularized feature utilization, producing a sparse and structured attribution profile

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 11 related papers). Coverage is partial, so this is neither replication nor refutation.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (38)

  1. Activation functions in deep learning: A comprehensive survey and benchmark (2022)
  2. Global prevalence of malnutrition: evidence from literature (2020)
  3. Handling Imbalanced Data in Insurance Fraud Detection: A Machine Learning Approach (2025)
  4. Automobile Insurance Fraud Detection Based on PSO-XGBoost Model and Interpretable Machine Learning Method (2025)
  5. The C omprehensive T ex A rchive N etwork ( CTAN ) (1993)
  6. R K H S - S H A P : S hapley values for kernel methods (2022)
  7. Bidding for the surplus: a non-cooperative approach to the S hapley value (2001)
  8. A non-cooperative mechanism for the Shapley value of airport problems (2015)
and 30 more
  1. Weightedshap: analyzing and improving shapley based feature attributions (2022)
  2. Instance S H A P : an instance-based estimation approach for S hapley values (2024)
  3. FW-Shapley: Real-Time Estimation of Weighted Shapley Values (2024)
  4. C H G S hapley: E fficient D ata V aluation and S election towards T rustworthy M achine L earning (2024) — link
  5. RankingSHAP–Listwise Feature Attribution Explanations for Ranking Models (2024) — link
  6. The value of cross-data set analysis for automobile insurance fraud detection (2022)
  7. Detecting insurance fraud using supervised and unsupervised machine learning (2023)
  8. Predicting automobile insurance fraud using classical and machine learning models (2024)
  9. A deep-learning–based antifraud system for car-insurance claims (2023)
  10. Automobile insurance fraud detection (2021)
  11. A study of model and hyper-parameter selection strategies for classifier ensembles: a robust analysis on different optimization algorithms and extended results (2021)
  12. Use of optimized Fuzzy C-Means clustering and supervised classifiers for automobile insurance fraud detection (2020)
  13. Vehicle Insurance Fraud Detection Based on Hybrid Approach for Data Augmentation. (2023)
  14. Hyperparameter optimization using genetic algorithms to detect frauds transactions (2021)
  15. Predicting fraud in financial payment services through optimized hyper-parameter-tuned XGBoost model (2022)
  16. Using machine learning models to compare various resampling methods in predicting insurance fraud (2021)
  17. Efficient resampling for fraud detection during anonymised credit card transactions with unbalanced datasets (2020)
  18. An Ensemble-Based Auto Insurance Fraud Detection Using BQANA Hyperparameter Tuning (2025)
  19. Focal loss for dense object detection (2017)
  20. Enhancing Auto Insurance Fraud Detection Using Convolutional Neural Networks (2024)
  21. Machine Learning based Method for Insurance Fraud Detection on Class Imbalance Datasets with Missing Values (2024)
  22. A convex surrogate operator for general non-modular loss functions (2016)
  23. CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition (2020)
  24. Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss (2019)
  25. Adafocal: Calibration-aware adaptive focal loss (2022)
  26. Calibrating deep neural networks using focal loss (2020)
  27. Statistical behavior and consistency of classification methods based on convex risk minimization (2004)
  28. Convex surrogates for unbiased loss functions in extreme classification with missing labels (2021)
  29. SMOTE: synthetic minority over-sampling technique (2002)
  30. The application of data mining techniques in financial fraud detection: A classification framework and an academic review of literature (2011)

0 of 38 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.