pathboost: A Python Package for Interpretable Graph-Level Prediction using Path-Based Gradient Boosting

PathBoost is an open-source Python package designed for supervised learning on graph-structured data. Unlike Graph Neural Networks (GNNs), which often function as black-box models, PathBoost provides an additive, interpretable model by explicitly identifying the labeled paths—sequences of node and edge attributes—that drive predictions. It is particularly well-suited for scientific domains like molecular chemistry, where understanding the structural substructures that influence a property is as important as the prediction itself.

The core of the package is a gradient boosting framework that avoids the computational burden of enumerating every possible path in a graph. Instead, it employs a "lazy" expansion strategy: it initializes with simple anchor nodes and iteratively selects the most predictive path at each step. This process separates path selection from path fitting, allowing users to use simple models (like decision stumps) for selection while using more complex base learners to fit the attribute-rich data associated with those paths. This modularity ensures that the model remains both computationally efficient and highly interpretable.

PathBoost is built to integrate seamlessly into existing data science workflows. It follows standard scikit-learn conventions, including support for cross-validation, and utilizes the networkx library for graph representation. Key methodological features include:

• Multi-anchor parallel architecture: Trains independent sub-models for different anchor labels to improve scalability.
• Flexible base learners: Supports custom estimators for fitting path-based features.
• Variable importance: Provides built-in tools to quantify the contribution of specific paths to the final prediction, including correlation adjustments to account for structural dependencies between nested paths.
• Task versatility: Supports both regression and binary classification, with specialized modes for better-calibrated probability estimation.