caplime: Conic Adaptive Piecewise LIME

caplime provides an educational R implementation of piecewise ridge regression with first- or second-degree hinge bases, and local surrogate explanations. The available local explanation modes are LIME and OPTILIME only. The main explanation function is CAP\_LIME().

Installation from the package folder

In RStudio, install devtools once if necessary. Then open the package project folder (the one containing DESCRIPTION) and run:

install.packages("devtools")
devtools::install("/path/to/caplime", dependencies = FALSE)
library(caplime)

Or build the CRAN source archive by running R CMD build caplime from the parent directory, and install the resulting caplime\_0.1.0.tar.gz.

Local explanations: LIME and OPTILIME

library(caplime)
set.seed(1)
x <- matrix(rnorm(300), ncol = 3)
colnames(x) <- c("a", "b", "c")
f <- function(newx) newx\[, 1] + newx\[, 2]^2

lime\_result <- CAP\_LIME(x, x\[1, ], f, method = "LIME", num\_samples = 200)
opti\_result <- CAP\_LIME(x, x\[1, ], f, method = "OPTILIME", num\_samples = 200)
print(head(lime\_result$effects))
print(head(opti\_result$effects))
plot\_CAP\_LIME(lime\_result)

Piecewise model: both degree settings

set.seed(2)
y <- x\[, 1]^2 + x\[, 2]
model\_degree1 <- fit\_CAP\_model(x, y, num\_knots = 2, degree = 1)
model\_degree2 <- fit\_CAP\_model(x, y, num\_knots = 2, degree = 2)
head(predict\_CAP\_model(model\_degree1, x))
head(predict\_CAP\_model(model\_degree2, x))

Testing

Run devtools::check("/path/to/caplime", cran = TRUE) in RStudio. The test scripts under tests/ exercise both degree = 1 and degree = 2, and both local explanation modes.

Scientific scope

The piecewise model uses ridge-penalized least squares with hinge basis functions, not an explicit conic solver. The OPTILIME mode searches a fixed set of bandwidth values to meet a local R-squared target; it is a simplified variant and does not reproduce the complete published OptiLIME procedure. CAP\_stability() returns a descriptive coefficient-variation table (CSI), not an inferential confidence-interval-based CSI.