h2oTreeHelpR

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

## Model Details:
## ==============
## 
## H2ORegressionModel: gbm
## Model ID:  GBM_model_R_1538296273962_40 
## Model Summary: 
##   number_of_trees number_of_internal_trees model_size_in_bytes min_depth
## 1               3                        3                 261         1
##   max_depth mean_depth min_leaves max_leaves mean_leaves
## 1         1    1.00000          2          2     2.00000
## 
## 
## H2ORegressionMetrics: gbm
## ** Reported on training data. **
## 
## MSE:  0.2193848
## RMSE:  0.4683854
## MAE:  0.4565917
## RMSLE:  0.3295538
## Mean Residual Deviance :  0.2193848

h2o_tree_dfs = h2o_tree_convertR(h2o_model = prostate.gbm)

h2o_tree_dfs[[1]]

##          node                                     node_text predictions
## 1 SG_0_Node_0 [shape=box, fontsize=14, label=GLEASON < 6.5]          NA
## 2 SG_0_Node_3             [fontsize=14, label=-0.019210527] -0.01921053
## 3 SG_0_Node_4              [fontsize=14, label=0.023479532]  0.02347953
##   node_text_label  left_split right_split left_split_levels
## 1   GLEASON < 6.5 SG_0_Node_3 SG_0_Node_4            [NA]|<
## 2    -0.019210527        <NA>        <NA>              <NA>
## 3     0.023479532        <NA>        <NA>              <NA>
##   right_split_levels NA_direction node_variable_type split_column
## 1                 >=         left            numeric      GLEASON
## 2               <NA>         <NA>               <NA>         <NA>
## 3               <NA>         <NA>               <NA>         <NA>
##   node_split_point
## 1              6.5
## 2               NA
## 3               NA

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]]

##   terminal_node terminal_node_depth      terminal_node_path
## 1   SG_0_Node_3                   1 SG_0_Node_0.SG_0_Node_3
## 2   SG_0_Node_4                   1 SG_0_Node_0.SG_0_Node_4
##           terminal_node_directions
## 1 (GLEASON < 6.5 | is.na(GLEASON))
## 2                   GLEASON >= 6.5

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]]

##   terminal_node terminal_node_depth      terminal_node_path
## 1   SG_0_Node_3                   1 SG_0_Node_0.SG_0_Node_3
## 2   SG_0_Node_4                   1 SG_0_Node_0.SG_0_Node_4
##           terminal_node_directions terminal_node_directions_h2o
## 1 (GLEASON < 6.5 | is.na(GLEASON))                            L
## 2                   GLEASON >= 6.5                            R

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