Which models in caret can use a sparse matrix for X?

Question

I would like to be able to use a sparse matrix as x in caret::train and it looks like many of them expect a data frame. I've been able to use sparse matrix with XGboost with caret but nnet and ELM both seem to require a data frame. I noticed in the code, caret tries to convert x to data frame for nnet and ELM models.

Is there a list of models that support sparse matrix?

Answer 1

You can use this piece of code to find which models are using as.matrix in the fit function.

Beware that as.matrix turns a sparse matrix into a full blown matrix. You might run into memory issues. I have not tested if the individual underlying models accept a sparse matrix.

library(caret)  # run on version 6.0-71

model_list <- getModelInfo()
df <- data.frame(models = names(model_list), 
                 fit = rep("", length(model_list)), 
                 stringsAsFactors = FALSE)

for (i in 1:length(model_list)) {
  df$fit[i] <- as.expression(functionBody(model_list[[i]]$fit))
}

# find xgboost matrix   
df$models[grep("xgb.DMatrix", df$fit)]
[1] "xgbLinear" "xgbTree"  

# find all models where fit contains as.matrix(x)
df$models[grep("as.matrix\\(x\\)", df$fit)]

[1] "bdk"               "binda"             "blasso"            "blassoAveraged"    "bridge"            "brnn"             
[7] "dnn"               "dwdLinear"         "dwdPoly"           "dwdRadial"         "enet"              "enpls.fs"         
[13] "enpls"             "foba"              "gaussprLinear"     "gaussprPoly"       "gaussprRadial"     "glmnet"           
[19] "knn"               "lars"              "lars2"             "lasso"             "logicBag"          "LogitBoost"       
[25] "lssvmLinear"       "lssvmPoly"         "lssvmRadial"       "mlpSGD"            "nnls"              "ordinalNet"       
[31] "ORFlog"            "ORFpls"            "ORFridge"          "ORFsvm"            "ownn"              "PenalizedLDA"     
[37] "ppr"               "qrnn"              "randomGLM"         "relaxo"            "ridge"             "rocc"             
[43] "rqlasso"           "rqnc"              "rvmLinear"         "rvmPoly"           "rvmRadial"         "sda"              
[49] "sddaLDA"           "sddaQDA"           "sdwd"              "snn"               "spikeslab"         "svmLinear"        
[55] "svmLinear2"        "svmLinear3"        "svmLinearWeights"  "svmLinearWeights2" "svmPoly"           "svmRadial"        
[61] "svmRadialCost"     "svmRadialSigma"    "svmRadialWeights"  "xgbLinear"         "xgbTree"           "xyf"