This article explores how stochastic regularization in neural networks can improve performance on unseen categorical data, especially high-cardinality categorical features. It uses visualizations and SHAP values to understand how entity embeddings respond to this regularization technique.