Gradient Boosted Trees on Complex Data Types

This project aims to bring advanced machine learning capabilities for handling sets and graphs into the mainstream gradient boosting frameworks XGBoost. The work is based on recent research from the Gilad-Bachrach group, including the Set-Tree and Tree-G algorithms, which extend decision trees and gradient-boosted trees to operate directly on complex data structures that commonly appear in medical and biological applications.

The Problem

Many real-world biomedical datasets contain information that is naturally represented as sets or graphs rather than fixed-length tables. Examples include:

• Sets of medications prescribed to a patient
• Collections of pre-existing medical conditions
• Protein interaction networks
• Contact tracing and biological interaction graphs

Traditional tree-based methods, despite being the leading approach for tabular data, cannot natively process such structures. As a result, researchers often rely on feature engineering techniques that may discard valuable information. Neural-network approaches can handle sets and graphs but typically underperform on tabular data and are often more difficult to deploy and interpret.

Impact and Market Opportunity

Decision-tree methods are among the most widely used machine learning techniques globally, with strong adoption in healthcare, bioinformatics, and industry. By integrating these capabilities into established open-source frameworks, the project could reach thousands of researchers and practitioners annually.

Potential applications include:

• Clinical outcome prediction
• Early disease detection
• Precision medicine
• Drug discovery
• General machine learning tasks involving structured data

Expected Outcome

The project will transform promising academic research into production-ready tools that can be adopted by the broader machine learning and biomedical research communities. By enabling gradient-boosted trees to operate directly on sets and graphs, it has the potential to significantly improve predictive modeling in healthcare and other domains while reducing reliance on complex feature engineering.

Related researches >>

Related videos >>

Gallery >>

Back to Bio-Computational Research lobby >>