Statistically testing training data for unwanted error patterns using rule-oriented regression
Year: 2025
Type: article
Abstract: • A method to detect biases in training data and de-poison it before training. • Examples of discovering biases and patterns in existing data sets. • Explainable AI via fuzzy reasoning combined with classical regression. • Using Boolean formulae to explain data, with statistical significance testing. • An open-source prototype implementation. Artificial intelligence (AI) models are only as good as the data they are trained on. Biases in training data can lead to biased outputs in machine learning models, a well-documented issue. However, methods to prevent these biases are less developed. Ensuring clean data during collection, such as using bias-aware sampling, is most effective when the entity collecting the data also trains the artificial intelligence (AI). When using pre-existing data, detecting if it has been manipulated or “poisoned” to induce undesired behavior in a model is challenging. We propose a method to test training data for flaws, establishing a trustworthy ground-truth for subsequent model training. Unlike methods that generate fuzzy rules from data, our approach defines rules before examining the data, allowing the discovery of hidden error patterns. Our method adds to statistical tests by inserting a flexible rule-checking phase that evaluates arbitrary user-defined, including fuzzy, data-pattern conditions beforehand. Those are then used as building blocks for regression models, using fuzzy inference inside. This enables explainability from fuzzy logic and statistical diagnostics from regression. Additionally, it is applicable to “small data,” not requiring large datasets like deep learning methods. We validate our method with an experimental open-source implementation, using real-life and synthetic datasets to ensure practical relevance and procedure validation.
Source: Expert Systems with Applications
Authors: Stefan Rass, Martin Dallinger
Institution: Johannes Kepler University of Linz
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,500
OpenAlex ID: W4416191332