Applicability domain based on ensemble learning in classification and regression analyses.

Area of Science:

• Machine Learning
• Cheminformatics
• Data Science

Background:

• Applicability domains (ADs) are crucial for assessing the reliability of predictive models in classification and regression.
• Existing AD methods in classification can be overly broad, limiting their practical utility.
• Ensemble learning offers potential for refining AD definitions.

Purpose of the Study:

• To propose a novel method for setting applicability domains (ADs) in classification and regression analyses.
• To enhance the reliability of predictions for new chemical compounds or data points.
• To address limitations of existing AD methods, particularly in classification tasks.

Main Methods:

• Developing an AD approach integrating ensemble learning with data density.
• Establishing a data density threshold to identify unreliable predictions.
• Applying ensemble learning for reliability assessment only on data exceeding the density threshold.
• Validating the proposed AD method using numerical simulations and quantitative structure-activity relationship (QSAR) data.

Main Results:

• The proposed method effectively sets appropriate ADs for both regression and classification analyses.
• Data density combined with ensemble learning provides a more refined and reliable AD than traditional methods.
• The approach successfully identifies regions where predictions are less trustworthy.

Conclusions:

• The integration of ensemble learning and data density offers a robust strategy for defining accurate applicability domains.
• This method enhances the trustworthiness of machine learning models in scientific applications.
• The validated approach provides a valuable tool for cheminformatics and related fields.