| Title: | LIME-Based Explanations with Interpretable Inputs Based on Ceteris Paribus Profiles |
| Version: | 0.5 |
| Maintainer: | Przemyslaw Biecek <przemyslaw.biecek@gmail.com> |
| Description: | Local explanations of machine learning models describe, how features contributed to a single prediction. This package implements an explanation method based on LIME (Local Interpretable Model-agnostic Explanations, see Tulio Ribeiro, Singh, Guestrin (2016) <doi:10.1145/2939672.2939778>) in which interpretable inputs are created based on local rather than global behaviour of each original feature. |
| URL: | https://github.com/ModelOriented/localModel |
| BugReports: | https://github.com/ModelOriented/localModel/issues |
| Depends: | R (≥ 3.5) |
| License: | GPL-2 | GPL-3 [expanded from: GPL] |
| Encoding: | UTF-8 |
| Imports: | glmnet, DALEX, ggplot2, partykit, ingredients |
| RoxygenNote: | 7.1.1 |
| Suggests: | covr, knitr, rmarkdown, randomForest, testthat |
| VignetteBuilder: | knitr |
| NeedsCompilation: | no |
| Packaged: | 2021-09-03 19:29:03 UTC; pbiecek |
| Author: | Przemyslaw Biecek [aut, cre], Mateusz Staniak [aut], Krystian Igras [ctb], Alicja Gosiewska [ctb], Harel Lustiger [ctb], Willy Tadema [ctb] |
| Repository: | CRAN |
| Date/Publication: | 2021-09-14 16:50:01 UTC |
LIME kernel from the original article with sigma = 1.
Description
Since only binary features are used, the weight associated with an observation is simply exp(-{number of features that were changed compared to the original observation}). Kernels are meant to be used as an argument to individual_surrogate_model function. Other custom functions can be used. Such functions take two vectors and return a single number.
Usage
gaussian_kernel(explained_instance, simulated_instance)
Arguments
explained_instance |
explained instance |
simulated_instance |
new observation |
Value
numeric
Examples
library(DALEX)
library(randomForest)
library(localModel)
data('apartments')
mrf <- randomForest(m2.price ~., data = apartments, ntree = 50)
explainer <- explain(model = mrf,
data = apartments[, -1])
model_lok <- individual_surrogate_model(explainer, apartments[5, -1],
size = 500, seed = 17,
kernel = gaussian_kernel)
# In this case each simulated observation has weight
# that is small when the distance from original observation is large,
# so closer observation have more weight.
model_lok
plot(model_lok)
LIME kernel that treats all observations as equally similar to the observation of interest.
Description
Kernels are meant to be used as an argument to individual_surrogate_model function. Other custom functions can be used. Such functions take two vectors and return a single number.
Usage
identity_kernel(explained_instance, simulated_instance)
Arguments
explained_instance |
explained instance |
simulated_instance |
new observation |
Value
numeric
Examples
library(DALEX)
library(randomForest)
library(localModel)
data('apartments')
mrf <- randomForest(m2.price ~., data = apartments, ntree = 50)
explainer <- explain(model = mrf,
data = apartments[, -1])
model_lok <- individual_surrogate_model(explainer, apartments[5, -1],
size = 500, seed = 17,
kernel = identity_kernel)
# In this case each simulated observation has equal weight
# when explanation model (LASSO) is fitted.
model_lok
plot(model_lok)
LIME-like explanations based on Ceteris Paribus curves
Description
This function fits a LIME-type explanation of a single prediction. Interpretable binary features that describe the local impact of features on the prediction are created based on Ceteris Paribus Profiles. Thend, a new dataset of similar observations is created and black box model predictions (scores in case of classification) are calculated for this dataset and LASSO regression model is fitted to them. This way, explanations are simplified and include only the most important features. More details about the methodology can be found in the vignettes.
Usage
individual_surrogate_model(
x,
new_observation,
size,
seed = NULL,
kernel = identity_kernel,
sampling = "uniform",
...
)
Arguments
x |
an explainer created with the function DALEX::explain(). |
new_observation |
an observation to be explained. Columns in should correspond to columns in the data argument to x. |
size |
number of similar observation to be sampled. |
seed |
If not NULL, seed will be set to this value for reproducibility. |
kernel |
Kernel function which will be used to weight simulated observations. |
sampling |
Parameter that controls sampling while creating new observations. |
... |
Additional arguments that will be passed to ingredients::ceteris_paribus. |
Value
data.frame of class local_surrogate_explainer
Examples
# Example based on apartments data from DALEX package.
library(DALEX)
library(randomForest)
library(localModel)
data('apartments')
mrf <- randomForest(m2.price ~., data = apartments, ntree = 50)
explainer <- explain(model = mrf,
data = apartments[, -1])
model_lok <- individual_surrogate_model(explainer, apartments[5, -1],
size = 500, seed = 17)
model_lok
plot(model_lok)
localModel: LIME-like explanations with interpretable features based on Ceteris Paribus profiles
Description
This package implements LIME-like explanation method (see Tulio Ribeiro, Singh, Guestrin (2016) <doi:10.1145/2939672.2939778>) in which interpretable inputs are created based on local rather than global behaviour of each original feature.#'
Important functions
individual_surrogate_model generates an explanation for a single prediction with
interpretable features based on Ceteris Paribus profiles.
plot.local_surrogate_explainer plots the explanation.
Generic plot function for local surrogate explainers
Description
Generic plot function for local surrogate explainers
Usage
## S3 method for class 'local_surrogate_explainer'
plot(x, ..., geom = "bar")
Arguments
x |
object of class local_surrogate_explainer |
... |
other objects of class local_surrogate_explainer. If provided, models will be plotted in rows, response levels in columns. |
geom |
If "point", lines with points at the end will be plotted, if "bar", bars will be plotted and if "arrow", arrows. |
Examples
# Example based on apartments data from DALEX package.
library(DALEX)
library(randomForest)
library(localModel)
data('apartments')
mrf <- randomForest(m2.price ~., data = apartments, ntree = 50)
explainer <- explain(model = mrf,
data = apartments[, -1])
model_lok <- individual_surrogate_model(explainer, apartments[5, -1],
size = 500, seed = 17)
model_lok
plot(model_lok)
Plot Ceteris Paribus Profile and discretization
Description
Plot Ceteris Paribus Profile and discretization
Usage
plot_interpretable_feature(x, variable)
Arguments
x |
local_surrogate_explainer object |
variable |
chr, name of the variable to be plotted |
Value
ggplot2 object
Generic print function for local surrogate explainers
Description
Generic print function for local surrogate explainers
Usage
## S3 method for class 'local_surrogate_explainer'
print(x, ...)
Arguments
x |
object of class local_surrogate_explainer |
... |
currently ignored |
Examples
# Example based on apartments data from DALEX package.
library(DALEX)
library(randomForest)
library(localModel)
data('apartments')
mrf <- randomForest(m2.price ~., data = apartments, ntree = 50)
explainer <- explain(model = mrf,
data = apartments[, -1])
model_lok <- individual_surrogate_model(explainer, apartments[5, -1],
size = 500, seed = 17)
plot(model_lok)
model_lok