Themen

Situation: The predictions created by machine learning models are increasingly used as decision support in organizations to reduce costs and increase profits. In high-stakes decision contexts, intrinsically interpretable machine learning (IIML) models are essential because they enable decision-makers to understand the model’s internal decision logic from data input to the prediction output and, ultimately, to trust its predictions. Problem: IIML algorithms learn patterns directly from the data they are trained on. Consequently, they may also learn spurious, incorrect, or misleading relationships present in the data. As a result, the intrinsic interpretations provided by these models, and communicated through graphical visualizations, may be distorted and inconsistent with established domain knowledge. Question: How to design a mechanism that aligns IIML models with domain knowledge? Solution: Development of a mechanism that aligns an IIML model with domain knowledge. For example, such a mechanism can be designed in two ways: i) as a post-hoc technique that adjusts model interpretations after model training or ii) as an integral component of the learning objective (e.g., the loss function) that adjusts model interpretations during model training. Literature: - Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206–215. - Mao, L., Wang, H., Hu, L. S., Tran, N. L., Canoll, P. D., Swanson, K. R., & Li, J. (2024). Knowledge-informed machine learning for cancer diagnosis and prognosis: A review. IEEE Transactions on Automation Science and Engineering. - Wang, Z. J., Kale, A., Nori, H., Stella, P., Nunnally, M., Chau, D. H., Vorvoreanu, M., Vaughan, J. W., & Caruana, R. (2021). GAM changer: Editing generalized additive models with interactive visualization. arXiv preprint arXiv:2112.03245.

Situation: The predictions created by machine learning models are increasingly used as decision support in organizations to reduce costs and increase profits. In high-stakes decision contexts, intrinsically interpretable machine learning (IIML) models are essential because they enable decision-makers to understand the model’s internal decision logic from data input to the prediction output and, ultimately, to trust its predictions. Problem: Most IIML algorithms learn only correlations between features (e.g., a higher number of higher ice cream sales is associated with a higher number of drowning incidents) to maximize performance in prediction tasks. In doing so, underlying causal relationships between features (e.g., hot weather increases both the number of ice cream sales and the number of drowning incidents) are often ignored. Consequently, the interpretations provided by these models, and visualized through plots, may be less meaningful for decision-makers because they reflect correlations and proxy variables rather than the true causal drivers of the outcome. Question: How to design a method for causal and intrinsically interpretable machine learning? Solution: Development of a method that combines causal machine learning with IIML. For example, a possible method can consist of three phases: i) discovering causal structures, ii) integrating the identified causal structures into an IIML model, and training the adapted model to create predictions and interpretations that are both transparent and causally grounded. Literature: - Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206–215. https://doi.org/10.1038/s42256-019-0048-x - Zapata Gonzalez, D. (2026). A step towards inherently interpretable causal machine learning models for decision support. In Proceedings of the 34th European Conference on Information Systems. AIS Electronic Library. https://aisel.aisnet.org/ecis2026/bus_analytics/bus_analytics/7 - Cui, P., & Athey, S. (2022). Stable learning establishes some common ground between causal inference and machine learning. Nature Machine Intelligence, 4(2), 110–115.