Robustness in Sum-Product Networks with continuous and categorical data

Sum-product networks are a popular family of probabilistic graphical models for which marginal inference can be performed in polynomial time. After learning sum-product networks from scarce data, small variations of parameters could lead to different conclusions. We adapt the robustness measure created for categorical credal sum-product networks to domains with both continuous and categorical variables. We apply this approach to a real-world dataset of online purchases where the goal is to identify fraudulent cases. We empirically show that such credal models can better discriminate between easy and hard instances than simply using the probability of the most probable class.