1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
|
% Generated by roxygen2: do not edit by hand
% Please edit documentation in R/xgb.plot.shap.R
\name{xgb.plot.shap}
\alias{xgb.plot.shap}
\title{SHAP contribution dependency plots}
\usage{
xgb.plot.shap(
data,
shap_contrib = NULL,
features = NULL,
top_n = 1,
model = NULL,
trees = NULL,
target_class = NULL,
approxcontrib = FALSE,
subsample = NULL,
n_col = 1,
col = rgb(0, 0, 1, 0.2),
pch = ".",
discrete_n_uniq = 5,
discrete_jitter = 0.01,
ylab = "SHAP",
plot_NA = TRUE,
col_NA = rgb(0.7, 0, 1, 0.6),
pch_NA = ".",
pos_NA = 1.07,
plot_loess = TRUE,
col_loess = 2,
span_loess = 0.5,
which = c("1d", "2d"),
plot = TRUE,
...
)
}
\arguments{
\item{data}{data as a \code{matrix} or \code{dgCMatrix}.}
\item{shap_contrib}{a matrix of SHAP contributions that was computed earlier for the above
\code{data}. When it is NULL, it is computed internally using \code{model} and \code{data}.}
\item{features}{a vector of either column indices or of feature names to plot. When it is NULL,
feature importance is calculated, and \code{top_n} high ranked features are taken.}
\item{top_n}{when \code{features} is NULL, top_n [1, 100] most important features in a model are taken.}
\item{model}{an \code{xgb.Booster} model. It has to be provided when either \code{shap_contrib}
or \code{features} is missing.}
\item{trees}{passed to \code{\link{xgb.importance}} when \code{features = NULL}.}
\item{target_class}{is only relevant for multiclass models. When it is set to a 0-based class index,
only SHAP contributions for that specific class are used.
If it is not set, SHAP importances are averaged over all classes.}
\item{approxcontrib}{passed to \code{\link{predict.xgb.Booster}} when \code{shap_contrib = NULL}.}
\item{subsample}{a random fraction of data points to use for plotting. When it is NULL,
it is set so that up to 100K data points are used.}
\item{n_col}{a number of columns in a grid of plots.}
\item{col}{color of the scatterplot markers.}
\item{pch}{scatterplot marker.}
\item{discrete_n_uniq}{a maximal number of unique values in a feature to consider it as discrete.}
\item{discrete_jitter}{an \code{amount} parameter of jitter added to discrete features' positions.}
\item{ylab}{a y-axis label in 1D plots.}
\item{plot_NA}{whether the contributions of cases with missing values should also be plotted.}
\item{col_NA}{a color of marker for missing value contributions.}
\item{pch_NA}{a marker type for NA values.}
\item{pos_NA}{a relative position of the x-location where NA values are shown:
\code{min(x) + (max(x) - min(x)) * pos_NA}.}
\item{plot_loess}{whether to plot loess-smoothed curves. The smoothing is only done for features with
more than 5 distinct values.}
\item{col_loess}{a color to use for the loess curves.}
\item{span_loess}{the \code{span} parameter in \code{\link[stats]{loess}}'s call.}
\item{which}{whether to do univariate or bivariate plotting. NOTE: only 1D is implemented so far.}
\item{plot}{whether a plot should be drawn. If FALSE, only a lits of matrices is returned.}
\item{...}{other parameters passed to \code{plot}.}
}
\value{
In addition to producing plots (when \code{plot=TRUE}), it silently returns a list of two matrices:
\itemize{
\item \code{data} the values of selected features;
\item \code{shap_contrib} the contributions of selected features.
}
}
\description{
Visualizing the SHAP feature contribution to prediction dependencies on feature value.
}
\details{
These scatterplots represent how SHAP feature contributions depend of feature values.
The similarity to partial dependency plots is that they also give an idea for how feature values
affect predictions. However, in partial dependency plots, we usually see marginal dependencies
of model prediction on feature value, while SHAP contribution dependency plots display the estimated
contributions of a feature to model prediction for each individual case.
When \code{plot_loess = TRUE} is set, feature values are rounded to 3 significant digits and
weighted LOESS is computed and plotted, where weights are the numbers of data points
at each rounded value.
Note: SHAP contributions are shown on the scale of model margin. E.g., for a logistic binomial objective,
the margin is prediction before a sigmoidal transform into probability-like values.
Also, since SHAP stands for "SHapley Additive exPlanation" (model prediction = sum of SHAP
contributions for all features + bias), depending on the objective used, transforming SHAP
contributions for a feature from the marginal to the prediction space is not necessarily
a meaningful thing to do.
}
\examples{
data(agaricus.train, package='xgboost')
data(agaricus.test, package='xgboost')
bst <- xgboost(agaricus.train$data, agaricus.train$label, nrounds = 50,
eta = 0.1, max_depth = 3, subsample = .5,
method = "hist", objective = "binary:logistic", nthread = 2, verbose = 0)
xgb.plot.shap(agaricus.test$data, model = bst, features = "odor=none")
contr <- predict(bst, agaricus.test$data, predcontrib = TRUE)
xgb.plot.shap(agaricus.test$data, contr, model = bst, top_n = 12, n_col = 3)
# multiclass example - plots for each class separately:
nclass <- 3
nrounds <- 20
x <- as.matrix(iris[, -5])
set.seed(123)
is.na(x[sample(nrow(x) * 4, 30)]) <- TRUE # introduce some missing values
mbst <- xgboost(data = x, label = as.numeric(iris$Species) - 1, nrounds = nrounds,
max_depth = 2, eta = 0.3, subsample = .5, nthread = 2,
objective = "multi:softprob", num_class = nclass, verbose = 0)
trees0 <- seq(from=0, by=nclass, length.out=nrounds)
col <- rgb(0, 0, 1, 0.5)
xgb.plot.shap(x, model = mbst, trees = trees0, target_class = 0, top_n = 4,
n_col = 2, col = col, pch = 16, pch_NA = 17)
xgb.plot.shap(x, model = mbst, trees = trees0 + 1, target_class = 1, top_n = 4,
n_col = 2, col = col, pch = 16, pch_NA = 17)
xgb.plot.shap(x, model = mbst, trees = trees0 + 2, target_class = 2, top_n = 4,
n_col = 2, col = col, pch = 16, pch_NA = 17)
}
\references{
Scott M. Lundberg, Su-In Lee, "A Unified Approach to Interpreting Model Predictions", NIPS Proceedings 2017, \url{https://arxiv.org/abs/1705.07874}
Scott M. Lundberg, Su-In Lee, "Consistent feature attribution for tree ensembles", \url{https://arxiv.org/abs/1706.06060}
}
|
