Multi-Source Causal Feature Selection.

Area of Science:

• Causal inference and machine learning
• Bioinformatics and computational biology
• Data science and artificial intelligence

Background:

• Causal feature selection is crucial for building reliable predictive models by identifying causal mechanisms.
• Existing methods typically analyze single datasets, limiting their application to multi-source data with varying distributions.
• There is a growing need for robust methods to perform causal feature selection on data from multiple sources, such as different gene expression datasets for the same disease.

Purpose of the Study:

• To develop a novel algorithm for causal feature selection that effectively handles data from multiple sources.
• To address the challenge of differing data distributions across datasets in multi-source causal feature selection.
• To leverage the principle of causal invariance to identify robust causal features.

Main Methods:

• Formulated multi-source causal feature selection as a search for an invariant set across datasets.
• Derived upper and lower bounds for the invariant set.
• Proposed the Multi-source Causal Feature Selection (MCFS) algorithm.

Main Results:

• MCFS effectively identifies causal features from multiple datasets with different distributions.
• Experimental validation using synthetic and real-world datasets demonstrated the superiority of MCFS.
• MCFS outperformed 16 other feature selection methods in identifying causal features.

Conclusions:

• MCFS provides a reliable approach for causal feature selection in multi-source settings.
• The algorithm's effectiveness is validated through extensive experiments on diverse datasets.
• MCFS advances the field of causal inference by enabling more robust model building from heterogeneous data sources.