Categorical features are a fundamental component of insurance data and are widely used in actuarial modeling. In this paper, we propose a novel knowledge-driven embedding method to learn numerical representations of categorical variables. The method consists of two key steps: First, we construct a graph to represent complex relationships and domain-specific insights associated with the categorical variables; in the second step, we employ state-of-the-art graph neural embedding techniques to transform the graph structures into meaningful numerical representations. A key advantage of this method is its ability to generate embeddings that not only represent each categorical level but encapsulates intrinsic patterns often overlooked by conventional embedding techniques, yielding richer and more informed representations. We illustrate the method’s practical application using a real-world automobile insurance dataset, showcasing its effectiveness in two critical scenarios: (1) generating robust risk classifications for high-cardinality categorical features, and (2) deriving reliable embeddings for new categories without prior claims experience. Our results demonstrate the method’s potential to enhance the predictive accuracy and improve the interpretability of actuarial models, ultimately contributing to more effective insurance pricing strategies. This work was supported by a 2024 CAS Individual Research Grant.
Kun Shi (Wed,) studied this question.