goat-behavior / calculate_shap.R
harpomaxx's picture
convert fastshap to dataframe
b98d704
raw
history blame
No virus
9.98 kB
suppressPackageStartupMessages(library(dplyr))
suppressPackageStartupMessages(library(fastshap)) # for fast (approximate) Shapley values
suppressPackageStartupMessages(library(caret))
#suppressPackageStartupMessages(library(doMC))
#registerDoMC(cores = 10)
p_function_G <-
function(object, newdata)
caret::predict.train(object, newdata = newdata, type = "prob")[, "G"]
p_function_GM <-
function(object, newdata)
caret::predict.train(object, newdata = newdata, type = "prob")[, "GM"]
p_function_R <-
function(object, newdata)
caret::predict.train(object, newdata = newdata, type = "prob")[, "R"]
p_function_W <-
function(object, newdata)
caret::predict.train(object, newdata = newdata, type = "prob")[, "W"]
# DEPRECATED
calculate_shap_deprecated <- function(dataset,model,nsim=10) {
# library(doParallel)
# registerDoParallel(8)
trainset <- dataset %>% na.omit() %>%
as.data.frame()
trainset_y <- dataset %>%
select(Activity) %>%
na.omit() %>%
unlist() %>%
unname()
trainset <- trainset %>% select(-Activity)
trainset_G <- trainset[which(trainset_y == "G"), ]
trainset_GM <- trainset[which(trainset_y == "GM"), ]
trainset_R <- trainset[which(trainset_y == "R"), ]
trainset_W <- trainset[which(trainset_y == "W"), ]
# Compute fast (approximate) Shapley values using 50 Monte Carlo repetitions
message(" - Calculating SHAP values for class G")
shap_values_G <-
fastshap::explain(
model,
X = trainset,
pred_wrapper = p_function_G,
nsim = nsim,
newdata = trainset_G
)
message(" - Calculating SHAP values for class GM")
shap_values_GM <-
fastshap::explain(
model,
X = trainset,
pred_wrapper = p_function_GM,
nsim = nsim,
newdata = trainset_GM
)
message(" - Calculating SHAP values for class R")
shap_values_R <-
fastshap::explain(
model,
X = trainset,
pred_wrapper = p_function_R,
nsim = nsim,
newdata = trainset_R
)
message(" - Calculating SHAP values for class W")
shap_values_W <-
fastshap::explain(
model,
X = trainset,
pred_wrapper = p_function_W,
nsim = nsim,
newdata = trainset_W
# adjust = TRUE
)
shap_values_GM$class<-"GM"
shap_values_G$class<-"G"
shap_values_R$class<-"R"
shap_values_W$class<-"W"
shap_values<-rbind(shap_values_G,
shap_values_GM,
shap_values_R,
shap_values_W)
shap_values
}
#' A new function for calcualting SHAP values
#' the function returns a dataframe with SHAP values in the same
#' order of the original dataset.
#'
#' SHAP value dataframe also contains information about Animal and
#' the prediction of the model. Notice that SHAP are calculated considering
#' the class (ground truth) and not the prediction. The prediction column is only
#' used for filtering ana analysis. The function `calculate_shapp_class()` can be
#' used for calculating SHAP values on prediction
#'
#' @param dataset a dataset used for calcuating SHAP. The dataset is used for
#' permutation during SHAP calculation and also each class is filtered and SHAP
#' value for each class is calculated.
#' @param model a model
#' @param nsim number of monte carlo simulation
#'
#' @return
#' @export
#'
#' @examples
calculate_shap <- function(dataset,model,nsim=10) {
trainset <- dataset %>% na.omit() %>%
as.data.frame()
trainset_y <- dataset %>%
select(Activity) %>%
na.omit() %>%
unlist() %>%
unname()
## Create an ID for maintaining the order
trainset <- cbind(id=seq(1:nrow(trainset)), trainset)
trainset <- trainset %>% select(-Activity)
trainset_G <- trainset[which(trainset_y == "G"), ]
trainset_GM <- trainset[which(trainset_y == "GM"), ]
trainset_R <- trainset[which(trainset_y == "R"), ]
trainset_W <- trainset[which(trainset_y == "W"), ]
id <- c(trainset_G$id,
trainset_GM$id,
trainset_R$id,
trainset_W$id)
trainset <- trainset %>% select(-id)
trainset_G <- trainset_G %>% select(-id)
trainset_GM <- trainset_GM %>% select(-id)
trainset_R <- trainset_R %>% select(-id)
trainset_W <- trainset_W %>% select(-id)
Anim <- c(trainset_G$Anim,
trainset_GM$Anim,
trainset_R$Anim,
trainset_W$Anim)
trainset <- trainset %>% select(-Anim)
trainset_G <- trainset_G %>% select(-Anim)
trainset_GM <- trainset_GM %>% select(-Anim)
trainset_R <- trainset_R %>% select(-Anim)
trainset_W <- trainset_W %>% select(-Anim)
predictions <- c(trainset_G$predictions,
trainset_GM$predictions,
trainset_R$predictions,
trainset_W$predictions)
trainset <- trainset %>% select(-predictions)
trainset_G <- trainset_G %>% select(-predictions)
trainset_GM <- trainset_GM %>% select(-predictions)
trainset_R <- trainset_R %>% select(-predictions)
trainset_W <- trainset_W %>% select(-predictions)
# Compute fast (approximate) Shapley values using 50 Monte Carlo repetitions
message(" - Calculating SHAP values for class G")
shap_values_G <-
fastshap::explain(
model,
X = trainset,
pred_wrapper = p_function_G,
nsim = nsim,
newdata = trainset_G
)
message(" - Calculating SHAP values for class GM")
shap_values_GM <-
fastshap::explain(
model,
X = trainset,
pred_wrapper = p_function_GM,
nsim = nsim,
newdata = trainset_GM
)
message(" - Calculating SHAP values for class R")
shap_values_R <-
fastshap::explain(
model,
X = trainset,
pred_wrapper = p_function_R,
nsim = nsim,
newdata = trainset_R
)
message(" - Calculating SHAP values for class W")
shap_values_W <-
fastshap::explain(
model,
X = trainset,
pred_wrapper = p_function_W,
nsim = nsim,
newdata = trainset_W
# adjust = TRUE
)
shap_values_G <- shap_values_G %>% as.data.frame()
shap_values_GM <- shap_values_GM %>% as.data.frame()
shap_values_R <- shap_values_R %>% as.data.frame()
shap_values_W <- shap_values_W %>% as.data.frame()
shap_values_G$class<-"G"
shap_values_GM$class<-"GM"
shap_values_R$class<-"R"
shap_values_W$class<-"W"
shap_values<-rbind(shap_values_G,
shap_values_GM,
shap_values_R,
shap_values_W)
shap_values <- shap_values %>% tibble::add_column(Anim)
shap_values <- shap_values %>% tibble::add_column(predictions)
#shap_values <-shap_values %>% tibble::add_column(id)
shap_values[order(id),]
}
#' Calculate SHAP values for a given PREDICTED class
#'
#' @param dataset the dataset used for permutation during SHAP calculation
#' @param new_data the new data we want to calculate SHAP
#' @param model the model used for explanation
#' @param nsim the number of Monte Carlos Simulations
#' @param function_class a wrapper function to obtain only a particular class
#' @param class_name the name of the class
#'
#' @return
#' @export
#'
#' @examples
#'
#' # Calculate the SHAP values for class G on new data
#' shap_values_G <- calculate_shap_class(
#' dataset,
#' new_data = newdata,
#' model = goat_model
#' nsim = 100,
#' function_class = p_function_G,
#' class_name = "G")
#'
#'
calculate_shap_class <- function(dataset, new_data, model,nsim=10,
function_class, class_name = "G") {
trainset <- dataset %>% na.omit() %>%
as.data.frame()
trainset_y <- dataset %>%
select(predictions) %>%
na.omit() %>%
unlist() %>%
unname()
trainset<- trainset %>%select (-Activity,-predictions,-Anim)
new_data_class <- new_data
Anim <- new_data_class$Anim
new_data_class <- new_data_class %>% select(-Anim)
Activity <- new_data_class$Activity
new_data_class <- new_data_class %>% select(-Activity)
predictions <- new_data_class$predictions
new_data_class <- new_data_class %>% select(-predictions)
# Compute fast (approximate) Shapley values using 50 Monte Carlo repetitions
message(" - Calculating SHAP values for class ",class_name)
shap_values_class <-
fastshap::explain(
model,
X = trainset,
pred_wrapper = function_class,
nsim = nsim,
newdata = new_data_class,
)
shap_values_class$class<-Activity
shap_values<-shap_values_class
shap_values <- shap_values %>% tibble::add_column(Anim)
shap_values <- shap_values %>% tibble::add_column(predictions)
shap_values
}
shap_summary_plot<-function(shap_values){
summary_plot <-
shap_values %>% reshape2::melt() %>% group_by(class, variable) %>%
summarise(mean = mean(abs(value))) %>%
arrange(desc(mean)) %>%
ggplot() +
ggdark::dark_theme_classic() +
geom_col(aes(
y = variable,
x = mean,
group = class,
fill = class
), position = "stack") +
xlab("Mean(|Shap Value|) Average impact on model output magnitude")
summary_plot
}
shap_beeswarm_plot<-function(shap_values,dataset){
shap_values <- shap_values %>% reshape2::melt()
dataset<-dataset %>% mutate(class=Activity) %>% select(-Activity) %>%
reshape2::melt() %>% group_by(variable) %>%
mutate(value_scale=range01(value))
beeswarm_plot<-cbind(shap_values, feature_value=dataset$value_scale) %>%
# filter(class=="GM") %>%
ggplot()+
facet_wrap(~class)+
#ggdark::dark_theme_bw()+
theme_classic()+
geom_hline(yintercept=0,
color = "red", size=0.5)+
ggforce::geom_sina(aes(x=variable,y=value,color=feature_value),size=0.5,bins=4,alpha=0.9,shape=15)+
scale_colour_gradient(low = "yellow", high = "red", na.value = NA)+
scale_colour_gradient(low = "skyblue", high = "orange", na.value = NA)+
xlab("Feature")+ylab("SHAP value")+
theme(axis.text.x = element_text(angle = 45, vjust = 0.5, hjust=1))
beeswarm_plot
}