predict.GMMCluster only returns cluster labels

[koheiw]

predict.GMMCluster <- function(object, newdata, ...) {
  predict_GMM(newdata, object$centroids, object$covariance_matrices, object$weights)$cluster_labels
}

ClusterR/R/clustering_functions.R

Line 200 in 02cef28

predict_GMM(newdata, object$centroids, object$covariance_matrices, object$weights)$cluster_labels

I wonder why this bug is not caught in the test:

ClusterR/tests/testthat/test-gmm.R

Lines 395 to 402 in 02cef28

	testthat::test_that("GMM predict method works", {
	gmm = GMM(dat, 3)
	testthat::expect_equal(predict_GMM(dat,
	CENTROIDS = gmm$centroids,
	COVARIANCE = gmm$covariance_matrices,
	WEIGHTS = gmm$weights)$cluster_labels,
	predict(gmm, dat))
	})

[mlampros]

@koheiw I'm sorry for the late reply.

I installed first the latest version from CRAN and then I used the documentation example and specifically the predict_GMM() which returns probabilities:

library(ClusterR)

data(dietary_survey_IBS)
str(dietary_survey_IBS)

dat = as.matrix(dietary_survey_IBS[, -ncol(dietary_survey_IBS)])

dat = center_scale(dat)
dat[1:10, 1:10]

gmm = GMM(data = dat, 
          gaussian_comps = 2, 
          dist_mode = "maha_dist", 
          seed_mode = "random_subset",
          km_iter = 10, 
          em_iter = 10, 
          verbose = TRUE)

pr = predict_GMM(data = dat, 
                 CENTROIDS = gmm$centroids, 
                 COVARIANCE = gmm$covariance_matrices, 
                 WEIGHTS = gmm$weights)
str(pr)
# List of 3
# $ log_likelihood: num [1:400, 1:2] -59.3 -61.7 -67.6 -67.9 -60.1 ...
# $ cluster_proba : num [1:400, 1:2] 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 ...
# $ cluster_labels: num [1:400] 1 1 1 1 1 1 1 1 1 1 ...

cluster_proba are the cluster probabilities.

[mlampros]

I now see your comment is related to the generic function

preds = predict(object = gmm, newdata = dat)
str(preds)
# num [1:400] 1 1 1 1 1 1 1 1 1 1 ...

You are right on this, and it requires that predict() returns the full list as the predict_GMM() function.
I'll fix this tomorrow and I'll push the changes.

[mlampros]

In first place, I took a look to the CRAN clustering packages and it seems that the majority of these packages return hard labels. A few exceptions are the following:

mclust::predict.Mclust()  # A list containing labels and posterior probabilitiesYes, as classification and z
clue::cl_predict()        # Hard IDs by defaultYes, through type = "memberships"

I then updated the functionality of the predict.GMMCluster function by adding the type argument which defaults to "cluster", i.e. the hard labels (the default remains backward compatible) and the user can receive probabilities using "prob" or the full list using "all"

predict.GMMCluster <- function(object, newdata,
                               type = c("cluster", "prob", "all"), ...) {

  type = match.arg(type)

  out = predict_GMM(newdata,
                    object$centroids,
                    object$covariance_matrices,
                    object$weights)

  switch(type,
         cluster = out$cluster_labels,
         prob = out$cluster_proba,
         all = out)
}

and I also updated the test case

library(ClusterR)

data(dietary_survey_IBS)
dat = as.matrix(dietary_survey_IBS[, -ncol(dietary_survey_IBS)])
dat = center_scale(dat)

gmm = GMM(data = dat, 
          gaussian_comps = 2, 
          dist_mode = "maha_dist", 
          seed_mode = "random_subset",
          km_iter = 10, 
          em_iter = 10, 
          verbose = TRUE)

expected = predict_GMM(dat,
                       CENTROIDS = gmm$centroids,
                       COVARIANCE = gmm$covariance_matrices,
                       WEIGHTS = gmm$weights)

testthat::expect_equal(predict(gmm, dat), expected$cluster_labels)
testthat::expect_equal(predict(gmm, dat, type = "cluster"), expected$cluster_labels)
testthat::expect_equal(predict(gmm, dat, type = "prob"), expected$cluster_proba)
testthat::expect_equal(predict(gmm, dat, type = "all"), expected)

I think the issue is now resolved and I'll proceed to close it. Feel free to re-open it if the code does not work as expected. I'll proceed in a couple of days to upload the updated version to CRAN.

Thanks for making me aware of this issue.