richardangell/gbmexplainr

Decompose gbm predictions into feature contributions

gbmexplainr's Issues

xgboost support

NA with factor variables

Example from decompose_gbm_prediction function results in the following error; Error in decompose_gbm_prediction(gbm1, data[1, ]) : ‘NA’ is not an expected level for X4

Add feature contribution verification

Add function to verify feature contributions against prediction from gbm.

Functionaility to plot individual trees

Plot single tree in model and show decision path for an observation through the tree.

LightGBM support

Input checking

• Check all columns required in model are in prediction data
• Check prediction data is data.frame and 1 row only
• Check categorical levels supplied are expected by model

No info in README

Add an example with pictures (output table + tree path) to README.md and #16

Function documentation

Complete documentation for all functions.

No option to specify subset of trees

No option to specify a subset of tree to use in the model.

Deal with link functions

Need to add functionality to deal with predictions returned on link scale (poisson + bernoulli distributions)

Error with ordered factors

Ordered factors have var.type = 0 so are compared as; value >= split point. This results in an NA as value is a factor level and _split point_is numeric. This then errors the surrounding if statement.

Results format

Output results in table with the following columns; variable, variable value, prediction contribution

Links to similar packages

Add links to other similar packages

Validation / reconciliation fail

Toy gbm from examples in help files fails the validate_decomposition function.

choose_split_r printing decision info

• Add option to supress printing of decision info
• Add data value used in split decision to printing

Add more information to split decision info

Add tree and node indexes to the split decision info returned.

No tests

Write tests for functions

CatBoost support

Overall package documentation

Add overall package documentation.

unknown error in choose_split_r

Sometimes when running the example from decompose_gbm_prediction function the following error occurs; Error in if (!num_value == pred_row[[split_col_name]]) { : missing value where TRUE/FALSE needed ....
Try re running the function a few times to replicate