Skip to content

edahelsinki/slise

Repository files navigation

SLISE - Sparse Linear Subset Explanations

Version Documentation Tests License: MIT

R implementation of the SLISE algorithm. The SLISE algorithm can be used for both robust regression and to explain outcomes from black box models. For more details see the conference paper, the robust regression paper or the local explanations paper. Alternatively for a more informal overview see the presentation, or the poster. Finally, there is also the documentation.

Björklund A., Henelius A., Oikarinen E., Kallonen K., Puolamäki K. (2019)
Sparse Robust Regression for Explaining Classifiers.
Discovery Science (DS 2019).
Lecture Notes in Computer Science, vol 11828, Springer.
https://doi.org/10.1007/978-3-030-33778-0_27

Björklund A., Henelius A., Oikarinen E., Kallonen K., Puolamäki K. (2022).
Robust regression via error tolerance.
Data Mining and Knowledge Discovery.
https://doi.org/10.1007/s10618-022-00819-2

Björklund A., Henelius A., Oikarinen E., Kallonen K., Puolamäki K. (2023)
Explaining any black box model using real data.
Frontiers in Computer Science 5:1143904.
https://doi.org/10.3389/fcomp.2023.1143904

The idea

In robust regression we fit regression models that can handle data that contains outliers (see the example below for why outliers are problematic for normal regression). SLISE accomplishes this by fitting a model such that the largest possible subset of the data items have an error less than a given value. All items with an error larger than that are considered potential outliers and do not affect the resulting model.

SLISE can also be used to provide local model-agnostic explanations for outcomes from black box models. To do this we replace the ground truth response vector with the predictions from the complex model. Furthermore, we force the model to fit a selected item (making the explanation local). This gives us a local approximation of the complex model with a simpler linear model (this is similar to, e.g., LIME and SHAP). In contrast to other methods SLISE creates explanations using real data (not some discretised and randomly sampled data) so we can be sure that all inputs are valid (follows the same constraints as when the data was generated, e.g., the laws of physics).

Installation

Using the devtools-package (install.packages("devtools")) install the slise package:

devtools::install_github("edahelsinki/slise")

After installation, load the package using:

library(slise)

Other Languages

The official Python version can be found here.

Example

In order to use SLISE you need to have your data in a numerical matrix (or something that can be cast to a matrix), and the response as a numerical vector. Below is an example of SLISE being used for robust regression:

library(slise)
library(ggplot2)
set.seed(42)

x <- seq(-1, 1, length.out = 50)
y <- -x + rnorm(50, 0, 0.15)
x <- c(x, rep(seq(1.6, 1.8, 0.1), 2))
y <- c(y, rep(c(1.8, 1.95), each = 3))

ols <- lm(y ~ x)$coefficients
slise <- slise.fit(x, y, epsilon = 0.5)

plot(slise, title = "", partial = TRUE, size = 2) +
    geom_abline(aes(intercept = ols[1], slope = ols[2], color = "OLS", linetype = "OLS"), size = 2) +
    scale_color_manual(values = c("#1b9e77", "#fda411"), name = NULL) +
    scale_linetype_manual(values = 2:1, name = NULL) +
    theme(axis.title.y = element_text(angle = 0, vjust = 0.5), legend.key.size = grid::unit(2, "line")) +
    guides(shape = FALSE, color = "legend", linetype = "legend")

Robust Regression Example Plot

SLISE can also be used to explain predictions from black box models such as convolutional neural networks:

library(slise)
set.seed(42)

source("experiments/explanations/data.R")
emnist <- data_emnist(digit=2)

slise <- slise.explain(emnist$X, emnist$Y, 0.5, emnist$X[17,], emnist$Y[17], logit=TRUE, lambda1=3, lambda2=6)
plot(slise, "image", "", c("not 2", "is 2"), plots = 1)

Explanation Example Plot

Dependencies

SLISE depends on the following R-packages:

  • Rcpp
  • RcppArmadillo
  • lbfgs

The following R-packages are optional, but needed for some of the built-in visualisations:

  • ggplot2
  • grid
  • gridExtra
  • reshape2
  • wordcloud