Universal feature selection for simultaneous interpretability of multitask datasets.

Area of Science:

• Computational chemistry
• Data science
• Machine learning

Background:

• Extracting meaningful features from high-dimensional scientific data is challenging due to scalability and restrictive assumptions.
• Existing methods often fail to capture complex feature-property interactions in large datasets.

Purpose of the Study:

• To develop a general and scalable feature selection algorithm (BoUTS) for complex, high-dimensional datasets.
• To identify both universal features applicable across datasets and task-specific features for subsets.
• To improve feature selection stability and enable knowledge transfer between scientific domains.

Main Methods:

• BoUTS employs a general and scalable feature selection algorithm.
• The algorithm identifies universal features relevant to all datasets and task-specific features for subsets.
• Evaluated on seven diverse chemical regression datasets.

Main Results:

• BoUTS achieves state-of-the-art feature sparsity while maintaining comparable prediction accuracy.
• Identified universal features enhance cross-dataset interpretability and knowledge transfer.
• Demonstrated scalability and applicability to tabular data across multiple domains.

Conclusions:

• BoUTS overcomes limitations of current feature selection methods by identifying universal and task-specific features.
• Universal features facilitate domain-specific knowledge transfer and provide insights into data-poor systems.
• The algorithm offers broad utility for manually-guided inverse problems and uncovering connections between chemical domains.