LocalGLMnet: A Deep Learning Architecture for Actuaries

The purpose of this talk is to discuss the LocalGLMnet architecture which is tailored to the needs of actuaries. The LocalGLMnet is a flexible network architecture for tabular data that allows for variable selection, the study of interactions, gives nice interpretations and allows to rank variable importance. The LocalGLMnet architecture is inspired by the structure and properties of generalized linear models (GLMs). It preserves the linear structure of GLMs, but it makes the coefficients of the linear predictors feature dependent. The LocalGLMnet architecture is similar to attention layers. Attention layers are a recently introduced new way of building powerful networks by extracting more important feature components from embeddings by giving more weight (attention) to them.

We exemplify the LocalGLMnet on a publicly available accident insurance data set. We show the nice properties and we have identified variables that can be dropped from the model.

This case study has been done as part of the "Data Science" working group of the Swiss Association of Actuaries (SAA). The group publishes tutorials that discuss the use of machine learning techniques for actuarial applications. The tutorials are self-explanatory and its code and data is publicly available on the website http://www.actuarialdatascience.org.