Mutual information-guided causal feature selection method for quality prediction of complex products

The multivariate nonlinear systems in complex product manufacturing pose challenges for quality modeling, where identifying key features is central to improving prediction accuracy and interpretability. However, most existing causal feature selection methods depend on conditional independence tests. These tests are costly and unstable in high dimensional scenarios, which motivates replacing them with computationally efficient and structurally parsimonious methods. This paper proposes a mutual information-guided causal feature selection method (MICFS) to construct the Markov boundary (MB) of target variables. First, a priority criterion is derived based on the data processing inequality. By comparing mutual information between features and the target, a parent-child variable set is preliminarily established. Second, a spouse variable criterion is proposed by integrating the activation principle of convergent structures. Heuristic rules are designed to jointly validate bidirectional screening results for the stability of conditional dependencies, forming a stable MB. Evaluations on benchmark Bayesian networks and industrial datasets show that MICFS reconstructs causal structures without conditional tests and outperforms baseline algorithms, offering a causally grounded feature subset for high dimensional quality prediction.