Causal Feature Selection With Dual Correction.

Area of Science:

• Causal inference
• Machine learning
• Feature selection

Background:

• Causal feature selection methods identify Markov boundaries (MB) using conditional independence (CI) tests.
• Real-world data issues like noise and small samples make CI tests unreliable, leading to false positives and false negatives in MB learning.
• Existing algorithms address only one type of error, failing to correct both simultaneously.

Purpose of the Study:

• To propose a dual-correction-strategy-based MB learning (DCMB) algorithm that addresses both false positives and false negatives simultaneously.
• To enhance DCMB with a simulated annealing approach (SA-DCMB) for automatic determination of optimal feature selection parameters.

Main Methods:

• Developed a dual-correction strategy to selectively remove false positives and retrieve false negatives in MB learning.
• Implemented a simulated annealing algorithm (SA-DCMB) to optimize the dual correction process.
• Evaluated performance on benchmark Bayesian network (BN) datasets and real-world datasets.

Main Results:

• DCMB demonstrated substantial improvements in MB learning accuracy compared to existing methods on BN datasets.
• SA-DCMB showed effectiveness in classification tasks on real-world datasets, outperforming state-of-the-art causal and traditional feature selection algorithms.

Conclusions:

• The proposed DCMB algorithm effectively corrects both false positives and false negatives in MB learning.
• SA-DCMB offers a robust and effective approach for causal feature selection in practical applications, enhancing classification accuracy.