Probabilistic Machine Learning Models for Risk-Sensitive Decision-Making

Abstract

Current ML based deterministic models fail to take into account the variability and risk during decision making under uncertain conditions. Currently used models aim at optimizing the expected performance but do not have the capability to account for risk. To this end, the current research proposes the use of a Stochastic Gradient-based Probabilistic Graphical Risk Network (SG-PGRNet) approach. The data analysis involves analyzing a structured dataset comprising of 4,000 observations and 17 numeric features that are intended to reveal uncertainty patterns and dynamic risk. In preprocessing, the data will be normalized using Z-scores to eliminate noise and ensure consistency; feature extraction using Independent Component Analysis (ICA) that involves deriving statistically independent components from a multivariate signal, and Stochastic Gradient-based Probabilistic Graphical Risk Network (SG-PGRNet) that utilizes probabilistic graphical models to derive relationships among variables to minimize expected risks during decision-making under uncertainty using Stochastic Gradient optimization. Results from applying the proposed model using Python are impressive in terms of its accuracy of 95.2%, precision of 93.5%, recall of 94.3%, F1-score of 94.9%, and AUC of 0.98. Compared to the conventional methods, it offers improved estimation of uncertainty and minimizes high-risk situations.