Learning Causal Fuzzy Logic Rules by Leveraging Markov Blankets
Te Zhang, Christian Wagner · 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC) · 2021
An important property of fuzzy systems is the interpretability provided by their rules. However, if a fuzzy system is derived through machine learning algorithms, its interpretability is often greatly diminished as membership functions and rules are adjusted to minimize error in respect to a data set. To address part of this challenge, we propose a novel two-step fuzzy rule generation framework leveraging the concept of the Markov blanket, i.e., the set of variables which are causally related to a target variable – such as a system’s output. By estimating the Markov blanket for a given application, we restrict rule learning to (only) the variables which are causally linked to the system output, thus minimising the generation of spurious rules (based on spurious correlations of variables). This decreases the complexity of fuzzy systems and maintains a causal link between rules’ antecedents and consequent(s) – as expected by humans when viewing rules. The proposed framework can improve the interpretability of fuzzy rule based systems which are tuned using machine learning techniques, while also providing performance advantages as are commonly associated with feature selection techniques. Experiment results show that even the initial implementation of the framework proposed here can generate more concise and interpretable rule bases without compromising performance.