Leveraging domain information to restructure biological prediction.

Area of Science:

• Machine Learning
• Data Science
• Computational Statistics

Background:

• Incorporating domain knowledge into prediction models is crucial but challenging.
• Discrete or categorical attributes can partition problem domains into simpler sub-problems.
• Identifying attributes that effectively simplify learning tasks is a key goal.

Purpose of the Study:

• Develop an algorithm to identify discrete/categorical attributes that maximally simplify supervised learning tasks.
• Propose a metric to rank attributes based on their potential to reduce classification uncertainty.
• Enhance prediction performance by restructuring learning problems.

Main Methods:

• Restructuring supervised learning problems using attribute-based partitions.
• Proposing a conditional entropy metric to quantify attribute effectiveness.
• Approximating solutions using expected minimum conditional entropy with random projections to manage computational cost.

Main Results:

• The proposed metric effectively ranks attributes for problem simplification.
• Tested on artificial, cheminformatics, and gene expression datasets.
• Classifiers built on restructured problems consistently outperform those on the original problem.

Conclusions:

• The conditional entropy metric successfully identifies effective partitions for classification problems.
• This approach enhances prediction performance by simplifying the learning task.
• The method provides a computationally feasible way to leverage domain knowledge encoded in discrete attributes.