Causally-Aware Unsupervised Feature Selection Learning.

Area of Science:

• Machine Learning
• Causal Inference
• Data Mining

Background:

• Unsupervised feature selection (UFS) is crucial for high-dimensional unlabeled data.
• Existing UFS methods often ignore causal relationships, leading to irrelevant features and poor interpretability.
• Graph-based UFS methods struggle with differentiating causal and non-causal features, creating inaccurate similarity graphs.

Purpose of the Study:

• To propose a novel UFS method, Causally-Aware UnSupErvised Feature Selection learning (CAUSE-FS), that addresses limitations of existing approaches.
• To enhance the interpretability and effectiveness of feature selection in unlabeled high-dimensional data by leveraging causal inference.
• To improve the construction of similarity graphs by accounting for the distinct roles of causal and non-causal features.

Main Methods:

• Introduced a causal regularizer to reweight samples, balancing confounding distributions for treatment features.
• Integrated the regularizer into a generalized unsupervised spectral regression model to reduce spurious feature-clustering associations.
• Employed causality-guided hierarchical clustering to group features by causal contribution and adaptively learned similarity graphs at multiple granularities.

Main Results:

• CAUSE-FS demonstrated superior performance compared to state-of-the-art UFS methods in extensive experiments.
• The method effectively mitigates spurious associations between features and clustering labels, achieving causal feature selection.
• Interpretability of the selected features was validated through visualization techniques.

Conclusions:

• CAUSE-FS offers a significant advancement in unsupervised feature selection by integrating causal inference.
• The proposed method enhances data analysis by improving feature relevance, interpretability, and the reliability of similarity graph construction.
• CAUSE-FS provides a robust framework for uncovering underlying causal structures in high-dimensional data.