Comprehensive decision tree models in bioinformatics.

Area of Science:

• Bioinformatics
• Machine Learning
• Data Mining

Background:

• Classification is a key machine learning technique in bioinformatics.
• There is a need for interpretable machine learning models for knowledge extraction and reasoning explanation.
• End-users often prefer comprehensible models over complex, black-box algorithms.

Purpose of the Study:

• To present an extension to a machine learning environment for visual tuning of decision tree classifiers.
• To develop an effective and easily interpretable decision tree model using a one-button data mining approach.
• To constrain model tuning by decision tree dimensions, avoiding bias from classification performance measures.

Main Methods:

• Developed a visual tuning method for decision tree classifiers.
• Integrated this method into an existing machine learning environment.
• Evaluated the approach on diverse datasets, including classical machine learning problems and bioinformatics data.

Main Results:

• Visually tuned decision trees showed a significant increase in accuracy, particularly for less complex models.
• Bioinformatics datasets exhibited higher accuracy gains compared to classical machine learning datasets.
• A user study confirmed significantly lower tuning times for the proposed visual tuning method versus manual tuning.

Conclusions:

• Building simple models with visual boundaries achieves high comprehensibility and classification performance.
• Visually tuned models perform comparably to more complex models built with default decision tree algorithms.
• The method is particularly suitable for bioinformatics datasets with binary attributes and numerous redundant features.