PathBoost Python Package Offers Interpretable Graph Prediction

A new Python package named `path_boost` has been released, providing a tool for interpretable supervised learning on graph-structured data. The package implements PathBoost, a gradient boosting algorithm designed to automatically identify predictive labeled paths within graphs during its learning process. This approach offers a significant advantage over many graph neural networks (GNNs), which often lack transparency in their decision-making. PathBoost generates an additive prediction model where path-based features explicitly reveal the specific substructures that influence predictions. To avoid the computationally intensive enumeration of all possible paths, the algorithm iteratively selects and extends paths based on their predictive power, combining these "weak learners" into a robust ensemble through boosting. The package is versatile, supporting both regression and binary classification tasks. Key features include seamless compatibility with scikit-learn workflows, the ability to use custom base learners and selectors, automatic starting node selection, and parallel training across anchor nodes. It also includes built-in variable importance computation, further enhancing interpretability. The utility of PathBoost was demonstrated on molecular property prediction and benchmarked against established GNNs and graph kernel methods across several molecular datasets.