MI-MCF: A Mutual Information-Based Multilabel Causal Feature Selection.

Area of Science:

• Machine Learning
• Causal Inference
• Data Mining

Background:

• Multilabel causal feature selection is crucial for understanding complex datasets.
• Existing methods often use computationally expensive conditional independence tests and do not differentiate label/feature contributions.

Purpose of the Study:

• To develop an efficient and effective multilabel causal feature selection algorithm.
• To address the limitations of existing Markov Blanket search methods in handling label-feature distinctness and computational cost.

Main Methods:

• Proposed the Mutual Information-based Multilabel Causal Feature Selection (MI-MCF) method.
• Utilized mutual information (MI) and conditional MI (CMI) to replace time-consuming conditional independence tests for Markov Blanket construction.
• Employed MI to weigh feature and label contributions to target nodes, aiding in feature recovery under label correlation.
• Implemented a symmetry check to eliminate spurious nodes.

Main Results:

• MI-MCF significantly reduces computational overhead compared to traditional methods.
• The method effectively identifies relevant features by considering unique label and feature contributions.
• MI-MCF autonomously determines the optimal number of features and shows superior performance on real-world datasets.

Conclusions:

• MI-MCF offers an efficient and accurate approach to multilabel causal feature selection.
• The method's ability to handle label correlation and distinct contributions enhances its practical applicability.
• Experimental results validate the effectiveness and superiority of MI-MCF over existing algorithms.