In richardangell/h2oTreeHelpR: h2o Tree Helper

Useful functions to manipulate tree based models in h2o from R

Refer to the Code-Structure document for an overview of the code structure.

The main functions in the package are demonstrated below.

First build a simple gbm with the following code;

library(h2o)
library(h2oTreeHelpR)
h2o.no_progress()
h2o.init()

prostate.hex = h2o.uploadFile(path = system.file("extdata",
                                                 "prostate.csv",
                                                 package = "h2o"),
                              destination_frame = "prostate.hex")
prostate.hex["RACE"] = as.factor(prostate.hex["RACE"])
prostate.hex["DPROS"] = as.factor(prostate.hex["DPROS"])
expl_cols <- c("AGE", "RACE", "DPROS", "DCAPS", "PSA", "VOL", "GLEASON")
prostate.gbm = h2o.gbm(x = expl_cols,
                       y = "CAPSULE",
                       training_frame = prostate.hex,
                       ntrees = 3,
                       max_depth = 1,
                       learn_rate = 0.1)

h2o_tree_convertR

Convert h2o tree structures to data.frames

prostate.gbm

h2o_tree_dfs = h2o_tree_convertR(h2o_model = prostate.gbm)

h2o_tree_dfs[[1]]

extract_split_rules

Get split conditions for terminal nodes for trees in a h2o model

terminal_node_rules <- extract_split_rules(h2o_tree_dfs)

terminal_node_rules[[1]]

map_h2o_encoding

Create a mapping for the output of h2o.predict_lead_node_assignment in terms of variable conditions

Specifically this a mapping between terminal node paths represented as i. L(eft) or R(ight) directions at each node (output from h2o.predict_lead_node_assignment) and ii. variable conditions (i.e. A > x & B in (y, z) etc).

terminal_node_mapping <- map_h2o_encoding(h2o_tree_dfs)

terminal_node_mapping[[1]]

The last column terminal_node_directions_h2o shows the terminal nodes as they are represented in the output of the h2o.predict_lead_node_assignment function.

richardangell/h2oTreeHelpR documentation built on July 3, 2019, 7:35 a.m.