Multi-View Causal Feature Selection.

Area of Science:

• Machine Learning
• Data Mining
• Causal Inference

Background:

• Multi-view feature selection leverages complementary information across datasets to reduce dimensionality and improve model performance.
• Existing methods often overlook causal relationships between features across views, limiting interpretability.
• Lack of causal understanding hinders the identification of truly influential features.

Purpose of the Study:

• To propose a novel Multi-View Causal Feature Selection (MVCFS) algorithm.
• To enhance feature selection interpretability by identifying causally relevant features across multiple data views.
• To improve classification accuracy through causally informed feature selection.

Main Methods:

• MVCFS algorithm uncovers causal relationships among features across different views using local causal structures.
• Constructs the complete Markov Blanket of the class variable in a multi-view setting.
• Employs causal inference techniques for feature selection.

Main Results:

• MVCFS identifies key features with causal interpretability across multiple views.
• Experimental results on 12 benchmark datasets show improved classification accuracy.
• Demonstrates superior performance compared to existing multi-view feature selection algorithms.

Conclusions:

• MVCFS offers a causally interpretable approach to multi-view feature selection.
• The algorithm effectively identifies important features by considering cross-view causal links.
• MVCFS represents a significant advancement in leveraging multi-view data for enhanced machine learning performance.