This package provides a goodness-of-fit test of whether a given i.i.d. sample {xi} is drawn from a given distribution. It works for any distribution once its score function (the derivative of log-density) ∇xlog p(x) can be provided. This method is based on ``A Kernelized Stein Discrepancy for Goodness-of-fit Tests and Model Evaluation’’ by Liu, Lee, and Jordan, available at <arXiv:1602.03253>.
The main function of this package is KSD, which estimates Kernelized Stein Discrepancy. Parameters include :
- x Sample of size Num_Instance x Num_Dimension
- score_function Score funtion (∇xlog p(x)) : takes x as input and output a column vector of size Num_Instance X Dimension. User may use pryr package to pass in a function that only takes in dataset as parameter, or user may also pass in computed score for a given dataset.
- kernel Type of kernel (default = ‘rbf’)
- width Bandwidth of the kernel
- nboot Bootstrap sample size
Other methods are also in this package, including various demos and examples.
KSD requires user to provide a score function to be used for computation. For example usage and exploration, a gmm class is provided in the package, which allow test KSD using gaussian mixture model.
Consider the following examples :
- We define a gmm, generate random data using the model, and test the null hypothesis that the data comes from the model. Obviously, the result will depend on the model and the amount of added random noise.
# Pass in a dataset generated by Gaussian distribution,
# pass in computed score rather than score function
library(KSD)
library(pryr)
model <- gmm()
X <- rgmm(model, n=100)
score_function = scorefunctiongmm(model=model, X=X)
result <- KSD(X,score_function=score_function)
result$p
#> [1] 0.899- We follow similar pattern, but in this example, we use pryr library to define a score function like a function handle in matlab, in which we pass in model as part of the function.
# Pass in a dataset generated by Gaussian distribution,
# use pryr package to pass in score function
library(KSD)
library(pryr)
model <- gmm()
X <- rgmm(model, n=100)
score_function = pryr::partial(scorefunctiongmm, model=model)
result <- KSD(X,score_function=score_function)
result$p
#> [1] 0.899Premade demos include the following (Note that these demos require additional libraries)
demo_iris()
demo_normal_performance()
demo_simple_gaussian()
demo_simple_gamma()
demo_gmm()
demo_gmm_multi()
A sample run of demo_iris :
library(KSD)
library(datasets)
library(ggplot2)
library(gridExtra)
library(mclust)
library(pryr)
demo_iris()
#> [1] "Fitting GMM with 3 clusters"
#> fitting ...
#> | | | 0% | |======== | 7% | |=============== | 13% | |======================= | 20% | |=============================== | 27% | |====================================== | 33% | |============================================== | 40% | |====================================================== | 47% | |============================================================= | 53% | |===================================================================== | 60% | |============================================================================= | 67% | |==================================================================================== | 73% | |============================================================================================ | 80% | |==================================================================================================== | 87% | |=========================================================================================================== | 93% | |===================================================================================================================| 100%
#> fitting ...
#> | | | 0% | |======== | 7% | |=============== | 13% | |======================= | 20% | |=============================== | 27% | |====================================== | 33% | |============================================== | 40% | |====================================================== | 47% | |============================================================= | 53% | |===================================================================== | 60% | |============================================================================= | 67% | |==================================================================================== | 73% | |============================================================================================ | 80% | |==================================================================================================== | 87% | |=========================================================================================================== | 93% | |===================================================================================================================| 100%
#> fitting ...
#> | | | 0% | |======== | 7% | |=============== | 13% | |======================= | 20% | |=============================== | 27% | |====================================== | 33% | |============================================== | 40% | |====================================================== | 47% | |============================================================= | 53% | |===================================================================== | 60% | |============================================================================= | 67% | |==================================================================================== | 73% | |============================================================================================ | 80% | |==================================================================================================== | 87% | |=========================================================================================================== | 93% | |===================================================================================================================| 100%
#> fitting ...
#> | | | 0% | |======== | 7% | |=============== | 13% | |======================= | 20% | |=============================== | 27% | |====================================== | 33% | |============================================== | 40% | |====================================================== | 47% | |============================================================= | 53% | |===================================================================== | 60% | |============================================================================= | 67% | |==================================================================================== | 73% | |============================================================================================ | 80% | |==================================================================================================== | 87% | |=========================================================================================================== | 93% | |===================================================================================================================| 100%
#> fitting ...
#> | | | 0% | |======== | 7% | |=============== | 13% | |======================= | 20% | |=============================== | 27% | |====================================== | 33% | |============================================== | 40% | |====================================================== | 47% | |============================================================= | 53% | |===================================================================== | 60% | |============================================================================= | 67% | |==================================================================================== | 73% | |============================================================================================ | 80% | |==================================================================================================== | 87% | |=========================================================================================================== | 93% | |===================================================================================================================| 100%
#> [1] "Average p value : 0.218"
Currently, the code is available at https://github.com/MinHyung-Kang/KSD/ More download options will be available after CRAN submission.
Minhyung(dot)Daniel(dot)Kang(at)gmail(dot)com