Probabilistic Graphical Models for Fraud and Anomaly Detection in Insurance
Details
An alternative title is
How to Build a Model with No Data and No Domain Knowledge...
Probabilistic Graphical Models (PGMs) are a way to think about incorporating structural dependencies across variables in a dataset. Bayesian Belief Networks are a class of PGM where the induced graph is a directed acyclic graph.
In this talk we will go through the basics of PGMs and Bayesian networks, and discuss how they might be used in insurance.
In particular we will discuss the use of Bayesian belief networks to help detect medical nondisclosure while underwriting life insurance.
If time permits, I will also briefly discuss potential uses for graphical models for anomaly and outlier detection where the data is categorical or unstructured.
Probabilistic Graphical Models for Fraud and Anomaly Detection in Insurance