Random Forest (RF) is an ensemble machine learning method that constructs many decision trees, each trained on a random subset of the data and features, and combines their predictions by majority vote (classification) or averaging (regression). It is widely used in earthquake engineering for fragility analysis, damage classification, and surrogate modeling.
Random Forest (RF), introduced by Leo Breiman in 2001, is an ensemble learning method that combines many individual decision trees to produce a single, more robust prediction. Each tree in the forest is trained on a bootstrap sample—a random sample drawn with replacement from the original dataset—and at each node, only a random subset of the input features is considered for splitting. This dual randomization (bootstrap sampling and feature subsampling) decorrelates the individual trees, which is the key to RF's remarkable robustness against overfitting.
The algorithm operates by constructing K decision trees independently, each growing on a different bootstrap sample. At each node of each tree, a random subset of m features (typically √M for classification or M/3 for regression, where M is the total number of features) is evaluated to find the best split. For classification tasks, the final prediction is determined by majority vote across all trees; for regression, the outputs are averaged. As the number of trees increases, the generalization error converges to a limit, meaning that adding more trees never causes overfitting—it only improves or stabilizes performance.
In earthquake engineering, RF has become one of the most widely used machine learning algorithms. Mangalathu and Jeon applied RF to seismic fragility analysis of bridge portfolios, and subsequent studies have used RF for post-earthquake damage assessment, tunnel vulnerability analysis, and high-rise shear wall fragility. A study on high-rise shear wall structures found that RF provided higher accuracy in seismic fragility analysis compared to multilayer perceptron (MLP) models. RF's ability to handle high-dimensional data, its robustness to noise, and its built-in feature importance estimation make it particularly suited for seismic applications where input parameters (PGA, magnitude, distance, structural properties) are numerous and their relationships are nonlinear.