[Paper Review] A Machine Learning-based Anomaly Detection Framework in Life Insurance Contracts

Andreas Groll, Akshat Khanna+1 more|arXiv (Cornell University)|Nov 26, 2024

Insurance and Financial Risk ManagementEconomics, Econometrics and Finance3 citations

TL;DR

This paper proposes a machine learning-based anomaly detection framework for life insurance contracts using unsupervised methods to identify data irregularities without labeled anomalies. It evaluates classical techniques like Isolation Forest and autoencoders on two open-source datasets, demonstrating that deep learning models—particularly variational autoencoders—achieve superior performance in detecting manually injected contract-level anomalies with high accuracy and automation potential.

ABSTRACT

Life insurance, like other forms of insurance, relies heavily on large volumes of data. The business model is based on an exchange where companies receive payments in return for the promise to provide coverage in case of an accident. Thus, trust in the integrity of the data stored in databases is crucial. One method to ensure data reliability is the automatic detection of anomalies. While this approach is highly useful, it is also challenging due to the scarcity of labeled data that distinguish between normal and anomalous contracts or inter\-actions. This manuscript discusses several classical and modern unsupervised anomaly detection methods and compares their performance across two different datasets. In order to facilitate the adoption of these methods by companies, this work also explores ways to automate the process, making it accessible even to non-data scientists.

Motivation & Objective

• To address the challenge of detecting anomalies in life insurance data where labeled anomalies are scarce or nonexistent.
• To compare the performance of classical unsupervised methods (e.g., k-means, DBSCAN, Isolation Forest, OCSVM) and modern deep learning models (autoencoders, VAEs) in identifying contract-level anomalies.
• To develop an automated, non-expert-friendly pipeline for anomaly detection to support real-world insurance operations.
• To evaluate model performance using manually injected anomalies to simulate realistic data irregularities in policy-level data.
• To provide a practical, scalable framework for insurers to enhance data integrity and fraud detection without requiring extensive data science expertise.

Proposed method

• Employs unsupervised anomaly detection techniques, including proximity-based methods (k-means, DBSCAN, HDBSCAN), tree-based Isolation Forest, and one-class SVM (OCSVM), to identify outliers in life insurance datasets.
• Applies deep learning models—autoencoders (AEs) and variational autoencoders (VAEs)—to learn low-dimensional representations and detect anomalies via reconstruction error.
• Uses grid search for automatic hyperparameter tuning in Isolation Forest (parameters: f_max, s_max, n_estimators), while other models require manual or experimental tuning.
• Preprocesses data using one-hot encoding for categorical variables and standardizes numerical features to ensure compatibility with all models.
• Evaluates model performance using a combination of metrics: number of detected anomalies out of four manually injected cases, reconstruction error (for AEs/VAEs), anomaly score (for Isolation Forest), and silhouette score (for clustering methods).
• Employs a threshold-based classification: anomalies are identified as instances with reconstruction error above a fixed threshold (0.5 for AEs, 0.7 for VAEs) or anomaly score above a learned cutoff.

Figure 1 : An example representation of anomalous and normal points in an example data set with two variables, $X$ and $Y$ , taken from Chandola et al. ( 2009 ) .

Experimental results

Research questions

• RQ1 How do classical unsupervised anomaly detection methods (e.g., Isolation Forest, OCSVM, k-means) perform in detecting contract-level anomalies in life insurance data with limited labeled data?
• RQ2 To what extent do deep autoencoder-based models (AEs and VAEs) outperform traditional machine learning methods in detecting anomalies in life insurance datasets?
• RQ3 Can automated hyperparameter tuning improve the robustness and reproducibility of anomaly detection models in a non-expert setting?
• RQ4 How do model runtime and scalability vary across different datasets, particularly when handling larger or more complex data structures?
• RQ5 What is the impact of architectural design (e.g., depth, latent dimension) on the performance and efficiency of autoencoders and variational autoencoders in anomaly detection tasks?

Key findings

• Variational autoencoders (VAEs) achieved the highest detection rate, identifying all four manually injected anomalies in both datasets, with 548 and 512 total anomalies detected on Dataset 1 and Dataset 2, respectively.
• Isolation Forest outperformed classical methods on both datasets, detecting all four anomalies on Dataset 1 and all four on Dataset 2, with a total of 342 and 1974 anomalies detected, respectively.
• Autoencoders detected 224 and 705 anomalies on Dataset 1 and Dataset 2, respectively, with 4 out of 4 anomalies correctly identified, and required less than 3 minutes of training time on Dataset 1.
• The best-performing VAE model on Dataset 2 used a learning rate of 1e-2, 950 training epochs, and a 10-dimensional latent space, achieving the highest anomaly detection recall.
• Classical methods like DBSCAN and HDBSCAN failed to complete on Dataset 2 due to high computational load, with runtime exceeding 5 hours, while Isolation Forest and OCSVM completed in under 10 seconds.
• The reconstruction error threshold of 0.7 for VAEs on Dataset 2 was optimal for maximizing true positive detection while minimizing false positives, as confirmed by manual inspection of results.

Figure 2 : Proportion of anomalies and detected manual anomalies (secondary $y$ -axis).