regression stats to shap.plot.dependence?

[arupakaa]

Dear Colleague,

You hard work is very much appreciated, as is your kind attention.

shap.plot.dependence is important validating the output of the SHAP model.
The histogram option is very useful for this, as is the smooth option.
However, the latter almost demands that the linear equation and R-squared be shown.

Would it be possible to implement as an option in shap.plot.dependence?

There is some easy code that enables this, found here:
https://gist.github.com/kdauria/524eade46135f6348140

And here is an example of it being used in practice, with output below:
library(devtools) source_gist("524eade46135f6348140") df = data.frame(x = c(1:100)) df$y = 2 + 5 * df$x + rnorm(100, sd = 40) df$class = rep(1:2,50) ggplot(data = df, aes(x = x, y = y, label=y)) + stat_smooth_func(geom="text",method="lm",hjust=0,parse=TRUE) + geom_smooth(method="lm",se=FALSE) + geom_point() + facet_wrap(~class)

Thanks in advance!

[liuyanguu]

Very cool. I will take a look! Many thanks

[liuyanguu]

Now in the latest version (0.0.4) there is an option to show Pearson correlation which is basically

plot1 <- plot1 + ggpubr::stat_cor(method = "pearson")

The default smooth line uses loess since most relationships are not linear. But you can make it linear by adding the geom_smooth yourself since the output is ggplot2 object (please turn add_hist off otherwise it is no longer ggplot2 object)

shap.plot.dependence(data_long = shap_long_iris, x="Petal.Length",
                     add_hist = FALSE, smooth = FALSE, add_stat_cor = TRUE) + 
  ggplot2::geom_smooth(method = 'lm', color = 'red', size = 0.4, se = F)

the default:

shap.plot.dependence(data_long = shap_long_iris, x="Petal.Length",
                     add_hist = TRUE, add_stat_cor = TRUE)

[arupakaa]

This is amazing. Thank you once again for the quick and responsive turn-around. Excellent progress being made all about. Cheers!