Establishing a Reproducible Feature-Engineering Baseline for Unsupervised First Notice of Loss Fraud Detection in Disability Claims Management Systems
DOI:
https://doi.org/10.63282/3050-9262.IJAIDSML-V2I3P111Keywords:
Unsupervised Anomaly Detection, Insurance Fraud, Feature Engineering, Disability Claims, ReproducibilityAbstract
Background: Insurance fraud in U.S. group benefits leaks tens of billions of dollars each year, but the academic literature on fraud detection still clusters around auto, health, and credit-card lines. Disability-claims fraud is missing from the peer-reviewed Scopus-indexed corpus. Production claims systems run on relational substrates typically IBM DB2 or Oracle yet the academic literature on in-database feature engineering for machine learning has been thin since MADlib in 2012. The recent Hopsworks Feature Store paper (2024) calls itself the first feature store at a top-tier database conference, which leaves a twelve-year gap in between. First Notice of Loss (FNOL) is the earliest signal in the claim lifecycle and the natural triage point, but it is industry vocabulary that rarely shows up in academic indexes. Three FNOL-exact peer-reviewed papers exist globally; none of them target disability claims. Methods: We present a portable PL/SQL feature schema that materialises eleven features at FNOL for disability claims: onset-date plausibility, last-employment-to-onset gap, occupational-class consistency, prior-claim density in the trailing twenty-four months, medical-provider claim rate, employer claim density, geo-temporal anomaly score, time-of-day deviation, claim-handle minutes, documentation completeness, and narrative-length z-score. The schema runs on IBM DB2 LUW and Oracle 19c with bitwise-identical output. We benchmark the schema against six unsupervised detectors via PyOD: Isolation Forest, Local Outlier Factor, One-Class SVM, COPOD, ECOD, and Robust Deep Autoencoder. Threshold calibration uses the ExCeeD per-example confidence method. The pipeline runs end-to-end on a synthetic disability-FNOL panel of 50,000 claims with 1.5 % contamination, generated from public actuarial schemas. Results: The feature schema is invariant across detector families. Cross-detector Spearman rank correlation on top 5 % suspicious claims is at least 0.71 across all detector pairs, with a maximum of 0.91 between COPOD and ECOD. Mean precision at the top 5 % converges within ±3.4 percentage points across the six detectors. ExCeeD-calibrated thresholds cut false-positive operational cost by 41.5 % on average compared to uncalibrated quantile thresholds at matched recall. DB2 and Oracle materialisations of the view layer produce feature values that match on every row to six decimal places; runtimes scale near-linearly to 5 million rows. Conclusions: Feature engineering, not detector choice, is the load-bearing reproducibility artifact for unsupervised disability-FNOL fraud detection. The schema is portable, requires no machine-learning middleware in the production claims path, and gives any future disability-fraud paper a citable baseline to extend. The reference implementation is released as supplementary material.
References
[1] F. T. Liu, K. M. Ting, and Z.-H. Zhou, "Isolation Forest," in Proc. ICDM 2008, 2008. DOI: 10.1109/ICDM.2008.17.
[2] M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, "LOF: Identifying Density-Based Local Outliers," in Proc. ACM SIGMOD 2000, 2000. DOI: 10.1145/342009.335388.
[3] B. Schölkopf, J. C. Platt, J. Shawe-Taylor, A. J. Smola, and R. C. Williamson, "Estimating the Support of a High-Dimensional Distribution," Neural Computation, vol. 13, no. 7, pp. 1443–1471, 2001. DOI: 10.1162/089976601750264965.
[4] Z. Li, Y. Zhao, N. Botta, C. Ionescu, and X. Hu, "COPOD: Copula-Based Outlier Detection," in Proc. ICDM 2020, 2020. DOI: 10.1109/ICDM50108.2020.00135.
[5] Liu, F. T., Ting, K. M., & Zhou, Z.-H. (2008). Isolation forest. In Proceedings of the 8th IEEE International Conference on Data Mining (pp. 413–422). IEEE. https://doi.org/10.1109/ICDM.2008.17
[6] C. Zhou and R. C. Paffenroth, "Anomaly Detection with Robust Deep Autoencoders," in Proc. ACM KDD 2017, 2017. DOI: 10.1145/3097983.3098052.
[7] Y. Zhao, Z. Nasrullah, and Z. Li, "PyOD: A Python Toolbox for Scalable Outlier Detection," Journal of Machine Learning Research, vol. 20, no. 96, pp. 1–7, 2019. arXiv: 1901.01588.
[8] Chandola, V., Banerjee, A., & Kumar, V. (2009). Anomaly detection: A survey. ACM Computing Surveys, 41(3), Article 15, 1–58. https://doi.org/10.1145/1541880.1541882
[9] L. Perini, V. Vercruyssen, and J. Davis, "Quantifying the Confidence of Anomaly Detectors in Their Example-Wise Predictions," in Proc. ECML PKDD 2020, LNAI 12459, pp. 227–243, 2020. DOI: 10.1007/978-3-030-67664-3_14.
[10] J. M. Hellerstein et al., "The MADlib Analytics Library or MAD Skills, the SQL," Proc. VLDB Endowment, vol. 5, no. 12, 2012. DOI: 10.14778/2367502.2367510.
[11] Ormenișan, A. A., Ismail, M., Hammar, K., Andersson, R., Gebremeskel, E., Kakantousis, T., Kouzoupis, A., Buso, F., Berthou, G., Dowling, J., & Haridi, S. (2019). Horizontally scalable ML pipelines with a feature store. In Proceedings of the 2nd SysML Conference
[12] D. Sculley et al., "Hidden Technical Debt in Machine Learning Systems," in Proc. NeurIPS 2015, 2015.
[13] B. J. Heil et al., "Reproducibility standards for machine learning in the life sciences," Nature Methods, vol. 18, pp. 1132–1135, 2021. DOI: 10.1038/s41592-021-01256-7.
[14] C. Phua, V. Lee, K. Smith, and R. Gayler, "A Comprehensive Survey of Data Mining-based Fraud Detection Research," arXiv: 1009.6119, 2010.
[15] Matloob, I., Khan, S. A., & Rahman, H. U. (2020). Sequence mining and prediction-based healthcare fraud detection methodology. IEEE Access, 8, 143256–143273. https://doi.org/10.1109/ACCESS.2020.3013962
[16] Subudhi, S., & Panigrahi, S. (2018). Detection of automobile insurance fraud using feature selection and data mining techniques. International Journal of Rough Sets and Data Analysis, 5(3), 1–20. https://doi.org/10.4018/IJRSDA.2018070101
[17] A. C. Bahnsen, D. Aouada, A. Stojanovic, and B. Ottersten, "Feature engineering strategies for credit card fraud detection," Expert Systems with Applications, vol. 51, pp. 134–142, 2016. DOI: 10.1016/j.eswa.2015.12.030.
[18] F. Carcillo, Y.-A. Le Borgne, O. Caelen, Y. Kessaci, F. Oblé, and G. Bontempi, "Combining unsupervised and supervised learning in credit card fraud detection," Information Sciences, vol. 557, pp. 317–331, 2021. DOI: 10.1016/j.ins.2019.05.042.
[19] A. E. Young, E. Besen, and J. Willetts, "The Relationship Between Work-Disability Duration and Claimant's Expected Time to Return to Work as Recorded by Workers' Compensation Claims Managers," Journal of Occupational Rehabilitation, vol. 27, no. 3, 2016. DOI: 10.1007/s10926-016-9656-z.
[20] Klambauer, G., Unterthiner, T., Mayr, A., & Hochreiter, S. (2017). Self-normalizing neural networks. Advances in Neural Information Processing Systems, 30. https://proceedings.neurips.cc/paper/2017/hash/5d44ee6f2c3f71a6d8c7b6f6b3f8b5e5-Abstract.html
[21] Polyzotis, N., Roy, S., Whang, S. E., & Zinkevich, M. (2017). Data management challenges in production machine learning. Proceedings of the 2017 ACM International Conference on Management of Data (SIGMOD ’17), 1723–1726. https://doi.org/10.1145/3035918.3054782
[22] Breck, E., Polyzotis, N., Roy, S., Whang, S. E., & Zinkevich, M. (2019). Data validation for machine learning. Proceedings of Machine Learning and Systems (SysML 2019).
