Introduction

Gradient Boosting Machines (GBMs) are powerful tree-based ensemble learning techniques used for regression and classification problems. They iteratively build decision trees, typically shallow, to improve the predictive performance by minimizing a loss function. Understanding feature interactions is essential in GBM models as it can reveal insights into the relationships between variables. Two popular R packages provide functions for exploring interactions in GBMs: `gbm::interact.gbm` and `dismo::gbm.interactions`. This article explores these functions, elucidates their technical aspects, compares their applications, and features an example to better illustrate their use.

Overview of gbm::interact.gbm

The `gbm::interact.gbm` function belongs to the `gbm` package and is tailored to assess interaction effects between predictors in the model. It computes the relative influence of interaction terms and provides estimates by fitting GBMs to resampled subsets of the data.

Key Features

• Purpose: Evaluate interaction effects between two predictors.
• Input: Takes a fitted `gbm` object, two variable indices, and optionally, a number of trees.
• Output: A numeric value representing the interaction strength.
• Usage: Suitable when you need a quick numerical assessment of specific two-way interactions in your model.

Example

• Purpose: Compute and visualize interaction strengths using Friedman's interaction statistic.
• Input: Requires a fitted `gbm` model object and optionally a threshold for interactions.
• Output: A list comprising interaction strengths and interaction plots.
• Usage: Ideal for a more comprehensive investigation of potential interactions across multiple predictors.
• Calculates the contribution of a pair or a set of variables to the joint prediction margin.
• Ranges from 0 (no interaction) to 1 (complete interaction).
• `gbm::interact.gbm`: Best suited for models where you suspect specific strong pairwise interactions.
• `dismo::gbm.interactions`: Preferred for mining and visualizing interactions among multiple variables for exploratory analysis.