Causal Feature Selection Algorithm Based on Maximizing Neighbourhood Mutual Information

The method of causal feature selection for constructing predictive models, which utilizes the causal relationships between predictor variables and class variables, has garnered significant attention in recent years. However, almost all existing causal feature selection algorithms use conditional independence (CI) tests to learn the Markov Blanket (MB), leading to algorithmic complexity that grows exponentially with the feature space and encountering errors in CI tests that affect model performance. In this paper, we analyze the unique role of neighborhood mutual information in causal relationships and propose a causal feature selection algorithm based on maximizing neighborhood mutual information (MNMI-CFS). Specifically, the MNMI-CFS method analyzes the relationships between features in three stages, selecting the features with the greatest causal impact on the target variable, thereby improving the performance of the predictive model. We validate the effectiveness of this method by comparing it with six advanced causal feature selection algorithms and demonstrate its performance advantages on five benchmark Bayesian networks and five real datasets.