XGBoost<\/code><\/a>,<\/li>LightGBM<\/code><\/a>,<\/li>fastshap<\/code><\/a>, and<\/li>treeshap<\/code><\/a>,<\/li><\/ul>\n\n\n\nsome of which with great visualizations. Plus there is SHAPforxgboost (see my recent post<\/a>), originally designed to visualize the results of SHAP values calculated from XGBoost, but it can also be used more generally by now.<\/p>\n\n\n\n
The shapviz package<\/h3>\n\n\n\n
![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/06\/logo.png\")
<\/figure>\n\n\n\nIn order to entangle calculation from visualization, the
shapviz<\/code> package was designed. It solely focuses on visualization of SHAP values. Closely following its README<\/a>, it currently provides these plots:<\/p>\n\n\n\nsv_waterfall()<\/code>: Waterfall plots to study single predictions.<\/li>sv_force()<\/code>: Force plots as an alternative to waterfall plots.<\/li>sv_importance()<\/code>: Importance plots (bar and\/or beeswarm plots) to study variable importance.<\/li>sv_dependence()<\/code>: Dependence plots to study feature effects (optionally colored by heuristically strongest interacting feature).<\/li><\/ul>\n\n\n\nThey require a “
shapviz<\/code>” object, which is built from two things only:<\/p>\n\n\n\nS<\/code>: Matrix of SHAP values<\/li>X<\/code>: Dataset with corresponding feature values<\/li><\/ol>\n\n\n\nFurthermore, a “
baseline<\/code>” can be passed to represent an average prediction on the scale of the SHAP values.<\/p>\n\n\n\nA key feature of the “
shapviz<\/code>” package is thatX<\/code> is used for visualization only. Thus it is perfectly fine to use factor variables, even if the underlying model would not accept these.<\/em><\/p>\n\n\n\nTo further simplify the use of
shapviz<\/code>, direct connectors to the packages<\/p>\n\n\n\nXGBoost<\/code><\/a>,<\/li>LightGBM<\/code><\/a>,<\/li>fastshap<\/code><\/a>, and<\/li>treeshap<\/code><\/a><\/li><\/ul>\n\n\n\nare available.<\/p>\n\n\n\n
Installation<\/h3>\n\n\n\n
The package
shapviz<\/code> can be installed from CRAN or Github:<\/p>\n\n\n\ndevtools::install_github(\"shapviz\")<\/code><\/li><\/ul>\n\n\n\ndevtools::install_github(\"mayer79\/shapviz\")<\/code><\/li><\/ul>\n\n\n\nExample<\/h3>\n\n\n\n
Shiny diamonds… let’s model their prices by four “c” variables with XGBoost, and create an explanation dataset with 2000 randomly picked diamonds.<\/p>\n\n\n
\n\t\t\t\n\t\t\t\t\n\t\t\t\tR<\/div>\n\t\t\t<\/div><\/div>\n\t\t\t<\/div>\n\t\t\t\n\nlibrary(shapviz)\nlibrary(ggplot2)\nlibrary(xgboost)\n\nset.seed(3653)\n\nX <- diamonds[c(\"carat\", \"cut\", \"color\", \"clarity\")]\ndtrain <- xgb.DMatrix(data.matrix(X), label = diamonds$price)\n\nfit <- xgb.train(\n params = list(learning_rate = 0.1, objective = \"reg:squarederror\"), \n data = dtrain,\n nrounds = 65L\n)\n\n# Explanation dataset\nX_small <- X[sample(nrow(X), 2000L), ]<\/pre><\/div>\n\n<\/div><\/div>\n\t\t<\/div>\n\n\n
Create “shapviz” object<\/h4>\n\n\n\n
One line of code creates a
shapviz<\/code> object. It contains SHAP values and feature values for the set of observations we are interested in. Note again thatX<\/code> is solely used as explanation dataset, not for calculating SHAP values.<\/p>\n\n\n\nIn this example we construct the
shapviz<\/code> object directly from the fitted XGBoost model. Thus we also need to pass a corresponding prediction datasetX_pred<\/code> used for calculating SHAP values by XGBoost.<\/p>\n\n\n\n\t\t\t\n\t\t\t\t\n\t\t\t\tR<\/div>\n\t\t\t<\/div><\/div>\n\t\t\t<\/div>\n\t\t\t\n\nshp <- shapviz(fit, X_pred = data.matrix(X_small), X = X_small)<\/pre><\/div>\n\n<\/div><\/div>\n\t\t<\/div>\n\n\n
Explaining one single prediction<\/h4>\n\n\n\n
Let’s start by explaining a single prediction by a waterfall plot or, alternatively, a force plot. <\/p>\n\n\n
\n\t\t\t\n\t\t\t\t\n\t\t\t\tR<\/div>\n\t\t\t<\/div><\/div>\n\t\t\t<\/div>\n\t\t\t\n\n# Two types of visualizations\nsv_waterfall(shp, row_id = 1)\nsv_force(shp, row_id = 1<\/pre><\/div>\n\n<\/div><\/div>\n\t\t<\/div>\n\n\n
![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/06\/README-waterfall.jpg\")
Waterfall plot<\/figcaption><\/figure>\n\n\n\n Factor\/character variables are kept as they are, even if the underlying XGBoost model required them to be integer encoded.<\/p>\n\n\n\n
![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/06\/README-force.jpg\")
Force plot<\/figcaption><\/figure>\n\n\n\n Explaining the model as a whole<\/h4>\n\n\n\n
We have decomposed 2000 predictions, not just one. This allows us to study variable importance<\/strong> at a global model level by studying average absolute SHAP values as a bar plot or by looking at beeswarm plots of SHAP values.<\/p>\n\n\n
\n\t\t\t\n\t\t\t\t\n\t\t\t\tR<\/div>\n\t\t\t<\/div><\/div>\n\t\t\t<\/div>\n\t\t\t\n\n# Three types of variable importance plots\nsv_importance(shp)\nsv_importance(shp, kind = \"bar\")\nsv_importance(shp, kind = \"both\", alpha = 0.2, width = 0.2)<\/pre><\/div>\n\n<\/div><\/div>\n\t\t<\/div>\n\n\n
![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/06\/README-imp1.png\")
Beeswarm plot<\/figcaption><\/figure>\n\n\n\n ![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/06\/README-imp2.jpg\")
Bar plot<\/figcaption><\/figure>\n\n\n\n ![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/06\/README-imp3.png\")
Beeswarm plot overlaid with bar plot<\/figcaption><\/figure>\n\n\n\n A scatterplot of SHAP values of a feature like
color<\/code> against its observed values gives a great impression on the feature effect<\/strong> on the response. Vertical scatter gives additional info on interaction effects.shapviz<\/code> offers a heuristic to pick another feature on the color scale with potential strongest interaction.<\/p>\n\n\n\n\t\t\t\n\t\t\t\t\n\t\t\t\tR<\/div>\n\t\t\t<\/div><\/div>\n\t\t\t<\/div>\n\t\t\t\n\nsv_dependence(shp, v = \"color\", \"auto\")<\/pre><\/div>\n\n<\/div><\/div>\n\t\t<\/div>\n\n\n
![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/06\/README-dep.png\")
Dependence plot with automatic interaction colorization<\/figcaption><\/figure>\n\n\n\n Summary<\/h3>\n\n\n\n
- The “shapviz” has a single purpose: making SHAP plots.<\/li>
- Its interface is optimized for existing SHAP crunching packages and can easily be used in future packages as well.<\/li>
- All plots are highly customizable. Furthermore, they are all written with
ggplot<\/code> and allow corresponding modifications.<\/li><\/ul>\n\n\n\n