shapviz

Introduction

SHAP (SHapley Additive exPlanations, see Lundberg and Lee (2017)) is an ingenious way to study black box models. SHAP values decompose - as fair as possible - predictions into additive feature contributions. Crunching SHAP values requires clever algorithms by clever people. Analyzing them, however, is super easy with the right visualizations. The “shapviz” package offers the latter:

These plots require a “shapviz” object, which is built from two things only:

  1. S: Matrix of SHAP values
  2. X: Dataset that includes the corresponding feature values

Furthermore, a baseline can be passed to represent an average prediction on the scale of the SHAP values.

A key feature of “shapviz” is that X is used for visualization only. Thus it is perfectly fine to use factor variables, even if the underlying model would not accept these. Additionally, in order to improve visualization, it can sometimes make sense to clip gross outliers, take logarithms for certain columns, or replace missing values by some explicit value.

To further simplify the use of “shapviz”, we added direct connectors to the R packages

For XGBoost, LightGBM, and H2O, the SHAP values are directly calculated from the fitted model.

CatBoost is not included, but see Section “Any other package” how to use its SHAP calculation backend with “shapviz”.

Installation

# From CRAN
install.packages("shapviz")

# Or the newest version from GitHub:
# install.packages("devtools")
devtools::install_github("mayer79/shapviz")

Example: Diamond prices

Fit model

We start by fitting an XGBoost model to predict diamond prices based on the four “C” features.

library(shapviz)
library(ggplot2)
library(xgboost)

set.seed(3653)

# Turn ordinal factors into normal ones
ord <- c("clarity", "cut", "color")
diamonds[, ord] <- lapply(diamonds[, ord], factor, ordered = FALSE)

# Fit XGBoost model
x <- c("carat", "clarity", "cut", "color")
dtrain <- xgb.DMatrix(data.matrix(diamonds[x]), label = diamonds$price)

fit <- xgb.train(
  params = list(learning_rate = 0.1, objective = "reg:squarederror"), 
  data = dtrain,
  nrounds = 65L
)

Create “shapviz” object

One line of code creates a “shapviz” object. It contains SHAP values and feature values for the set of observations we are interested in. Note again that X is solely used as explanation dataset, not for calculating SHAP values.

In this example we construct the “shapviz” object directly from the fitted XGBoost model. Thus we also need to pass a corresponding prediction dataset X_pred used for calculating SHAP values by XGBoost.

# Pick explanation data
dia_small <- diamonds[sample(nrow(diamonds), 2000L), ]

# We also pass feature data X with originally encoded values
shp <- shapviz(fit, X_pred = data.matrix(dia_small[x]), X = dia_small)

Note: If X_pred would contain one-hot-encoded dummy variables, their SHAP values could be collapsed by the collapse argument of shapviz().

Decompose single prediction

The main idea behind SHAP values is to decompose, in a fair way, a prediction into additive contributions of each feature. Typical visualizations include waterfall plots and force plots:

sv_waterfall(shp, row_id = 1L) +
  theme(axis.text = element_text(size = 11))

Works pretty sweet, and factor input is respected!

Alternatively, we can study a force plot:

sv_force(shp, row_id = 1L)

SHAP importance

Studying SHAP decompositions of many observations allows to gain an impression on variable importance. As simple descriptive measure, the mean absolute SHAP value of each feature is considered. These values can be plotted as a simple bar plot, or, to add information on the sign of the feature effects, as a beeswarm plot sorted by the mean absolute SHAP values. Such beeswarm plots are often called “summary plots”.

# A bar plot of mean absolute SHAP values
sv_importance(shp)


# A beeswarm plot
sv_importance(shp, kind = "beeswarm")


# Or both!
sv_importance(shp, kind = "both", show_numbers = TRUE, width = 0.2)

SHAP dependence plots

A SHAP beeswarm importance plot gives first hints on whether high feature values tend to high or low predictions. This impression can be substantiated by studying simple scatterplots of SHAP values of a feature against its feature values. A second feature can be added as color information to see whether the feature effect depends on the feature on the color scale or not. The stronger the vertical scatter for similar values on the x axis, the stronger the interactions.

sv_dependence(shp, v = "color", color_var = "auto")


sv_dependence(shp, v = "carat", color_var = "auto", alpha = 0.2, size = 1) +
  guides(colour = guide_legend(override.aes = list(alpha = 1, size = 2)))

Interface to other packages

The above example uses XGBoost to calculate SHAP values. In the following sections, we show (without running the code), how other packages work together with “shapviz”.

LightGBM

library(lightgbm)
dtrain <- lgb.Dataset(data.matrix(diamonds[x]), label = diamonds$price)

fit <- lgb.train(
  params = list(learning_rate = 0.1, objective = "regression"), 
  data = dtrain,
  nrounds = 65L
)

shp <- shapviz(fit, X_pred = data.matrix(dia_small[x]), X = dia_small)
sv_importance(shp)

fastshap

library(fastshap)

# SHAP sampling values using iris data
fit <- lm(Sepal.Length ~ Sepal.Width + Petal.Length, data = iris)

shap <- explain(
  fit, 
  X = iris[c("Sepal.Width", "Petal.Length")], 
  newdata = iris[c("Sepal.Width", "Petal.Length")], 
  nsim = 100, 
  pred_wrapper = predict
)

sv <- shapviz(shap, X = iris)
sv_dependence(sv, "Sepal.Width")

shapr

library(shapr)

fit <- lm(price ~ carat + clarity + cut + color, data = diamonds)
dia_smaller <- diamonds[sample(nrow(diamonds), 20L), ]

explainer <- shapr(dia_smaller, fit)

explanation <- explain(
  dia_smaller,
  approach = "ctree",
  explainer = explainer,
  prediction_zero = mean(diamonds$price)
)

shp <- shapviz(explanation)
sv_dependence(shp, "carat", "auto")

H2O

If you work with a tree-based H2O model:

library(h2o)

h2o.init()

dia_h2o <- as.h2o(diamonds)

fit <- h2o.gbm(x, "price", training_frame = dia_h2o)
shp <- shapviz(fit, X_pred = dia_small)
sv_force(shp, row_id = 1)
sv_dependence(shp, "carat", "auto")

treeshap

library(treeshap)
library(ranger)

fit <- ranger(
  y = diamonds$price, x = data.matrix(diamonds[x]), max.depth = 4, num.trees = 100
)
unified_model <- ranger.unify(fit, data.matrix(diamonds[x]))
shaps <- treeshap(unified_model, data.matrix(dia_small[x]))
shp <- shapviz(shaps, X = dia_small)
sv_importance(shp)
sv_dependence(shp, "clarity", color_var = "auto", alpha = 0.2, size = 1)

kernelshap

library(kernelshap)

dia_smaller <- diamonds[sample(nrow(diamonds), 100L), ]
fit <- lm(price ~ carat + clarity + cut + color, data = diamonds)
ks <- kernelshap(fit, dia_small[x], bg_X = dia_smaller)

shp <- shapviz(ks)
sv_importance(shp)
sv_dependence(shp, "carat")

Any other package

The most general interface is to provide a matrix of SHAP values and corresponding feature values (and optionally, a baseline value):

S <- matrix(c(1, -1, -1, 1), ncol = 2, dimnames = list(NULL, c("x", "y")))
X <- data.frame(x = c("a", "b"), y = c(100, 10))
shp <- shapviz(S, X, baseline = 4)

An example is CatBoost: it is not on CRAN and requires catboost.*() functions to calculate SHAP values, so we cannot directly add it to “shapviz” for now. Just use a wrapper like this:

library(catboost)

shapviz.catboost.Model <- function(object, X_pred, X = X_pred, collapse = NULL, ...) {
  if (!requireNamespace("catboost", quietly = TRUE)) {
    stop("Package 'catboost' not installed")
  }
  stopifnot(
    "X must be a matrix or data.frame. It can't be an object of class catboost.Pool" =
      is.matrix(X) || is.data.frame(X),
    "X_pred must be a matrix, a data.frame, or a catboost.Pool" =
      is.matrix(X_pred) || is.data.frame(X_pred) || inherits(X_pred, "catboost.Pool"),
    "X_pred must have column names" = !is.null(colnames(X_pred))
  )
  
  if (!inherits(X_pred, "catboost.Pool")) {
    X_pred <- catboost.load_pool(X_pred)
  }

  S <- catboost.get_feature_importance(object, X_pred, type = "ShapValues", ...)

  # Call matrix method
  pp <- ncol(X_pred) + 1L
  baseline <- S[1L, pp]
  S <- S[, -pp, drop = FALSE]
  colnames(S) <- colnames(X_pred)
  shapviz(S, X = X, baseline = baseline, collapse = collapse)
}

# Example
X_pool <- catboost.load_pool(diamonds[x], label = diamonds$price)

fit <- catboost.train(
  X_pool, 
  params = list(
    loss_function = "RMSE", 
    iterations = 165, 
    logging_level = "Silent", 
    allow_writing_files = FALSE
  )
)

shp <- shapviz(fit, X_pred = dia_small[x])
sv_importance(shp)
sv_dependence(shp, "clarity", color_var = "auto", alpha = 0.2, size = 1)

Classification models

The plot functions work with one-dimensional model predictions only. However, the wrappers for XGBoost, LightGBM, and kernelshap allow to select the category of interest.

References

Lundberg, Scott M, and Su-In Lee. 2017. “A Unified Approach to Interpreting Model Predictions.” In Advances in Neural Information Processing Systems 30, edited by I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, 4765–74. Curran Associates, Inc. https://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions.pdf.