R Implementation of Bayesian Additive Semi-Structured Regression Trees (BAMDT).

Reference:

Luo, Z. T., Sang, H., & Mallick, B. (2022) BAMDT: Bayesian Additive Semi-Multivariate Decision Trees for Nonparametric Regression. Proceedings of the 39th International Conference on Machine Learning (ICML 2022) [link]

• Demo.R: Demo code for fitting BAMDT on a U-shape domain
• Tree.R: Class of semi-structured decision trees
• Model.R: Class of BAMDT models
• ComplexDomainFun.R: Utility functions for U-shape domain
• SimData.R: Code to simulate data on a U-shape domain (no need to run unless you would like to regenerate data)
• input_U.RData: Data sets generated by running SimData.R

The code depends on the following R packages: R6, collections, igraph, fdaPDE, BART, sf, ggplot2. Please make sure they are installed before running the demo code.

model = Model$new(Y, X, graphs, projections, hyperpar, X_new, projections_new)

creates a BAMDT model object named model.

Parameters:

• Y: Numeric responses vector of length n.
• X: Numeric unstructured training features of size n * p.
• graphs: List of M spatial graphs, where M is the number of trees. Each graph should be an igraph object.
• projections: Integer matrix of size n * M, where projections[i, j] is the nearest knot index corresponding to training observation i for tree j.
• hyperpar: Named vector of hyperparameters with the following elements
  • hyperpar['M']: Number of trees M.
  • hyperpar['sigmasq_mu']: Variance of prior for $\mu$, i.e., $\sigma^2_\mu$.
  • hyperpar['q']: Quantile used to calibrate prior for noise variance $\sigma^2$.
  • hyperpar['nu']: Degree of freedom of the inverse-$\chi^2$ prior for noise variance $\sigma^2$.
  • hyperpar['alpha']: Hyperparameter $\alpha$ in tree generating process.
  • hyperpar['beta']: Hyperparameter $\beta$ in tree generating process.
  • hyperpar['numcut']: Number of candidate split points for unstructured features.
  • hyperpar['prob_split_by_x']: Probability for performing a unstructured split.
• X_new: Numeric unstructured test features of size n_new * p.
• projections_new: Integer matrix of size n_ho * M, where projections_new[i, j] is the nearest knot index corresponding to test observation i for tree j.

To fit a BAMDT model and predict for test data, use

model$Fit(init_val, MCMC, BURNIN, THIN, seed = 1234, save_partitions = FALSE)

Parameters:

• init_val: Named list of initial values with the following element
  • init_val[['sigmasq_y']]: Initial value for noise variance $\sigma^2$.
• MCMC: Number of MCMC iterations.
• BURNIN: Number of burn-in iterations.
• THIN: Retain MCMC samples every THIN iterations, i.e., the number of posterior samples is npost = (MCMC - BURNIN) / THIN.
• seed: Random seed.
• save_partition: Logical value indicating whether posterior samples of partitions are saved. Default is FALSE (recommended). Setting save_partition = TRUE is highly memory inefficient.

The model object has the following public members:

• model$sigmasq_y_out: Posterior samples of noise variance $\sigma^2$.
• model$g_out: npost * n * M array of posterior samples of (in-sample) fitted values from each tree.
• model$Y_new_out: npost * n_new matrix of posterior samples of (out-of-sample) predicted values.
• model$importance_out: npost * (p + 1) matrix of posterior samples of feature importance metrics.