Counterfactual Uncertainty for High Dimensional Tabular Dataset

With the advent of machine learning (ML) and deep learning in geophysics interpretation tasks, ML models especially classifiers have proven valuable in assessing hydrocarbon exploration risks using tabular data. However, these models are often considered”black-box” due to their lack of interpretability. To address this, explainable artificial intelligence (XAI) methods are employed, in an attempt to provide insights to the inner workings and behaviors of these models. Our research introduces a workflow empowering users to comprehend these black-box classifiers through counterfactual generation. By perturbing features in tabular datasets, novel instances called Forward Counterfactuals are crafted and are assigned uncertainty scores, in order to ground their validity. Inference testing on these generated datapoints reveals valuable insights into model responses, enhancing decision-making, fairness analysis, and understanding of influencing factors. Our paper provides a case study of the usage of this setup for a hydrocarbon prospect risking dataset, however it offers promising contributions for broader ML applications in diverse domains.