Production mode • compboost

Storing the complete [Compboost] object requires to save a lot of data:

Data matrices of the raw data.
Transformed data matrices. Each base learner creates a design matrix with (potentially) multiple columns.

Hence, compboost allows to store the model without the data. Within this vignettes, this is also called production mode since it is the more practical case when running the model in production.

Store model without data

To do so, just call:

dat = mlr3::tsk("sonar")$data()
cboost = boostSplines(dat, "Class", oob_fraction = 0.3)
#>   1/100   risk = 0.68  oob_risk = 0.68   time = 0   
#>   2/100   risk = 0.67  oob_risk = 0.67   time = 1007   
#>   4/100   risk = 0.66  oob_risk = 0.67   time = 2819   
#>   6/100   risk = 0.66  oob_risk = 0.66   time = 4621   
#>   8/100   risk = 0.65  oob_risk = 0.66   time = 6438   
#>  10/100   risk = 0.65  oob_risk = 0.65   time = 8323   
#>  12/100   risk = 0.64  oob_risk = 0.65   time = 10197   
#>  14/100   risk = 0.63  oob_risk = 0.64   time = 12052   
#>  16/100   risk = 0.63  oob_risk = 0.64   time = 13910   
#>  18/100   risk = 0.62  oob_risk = 0.64   time = 15762   
#>  20/100   risk = 0.62  oob_risk = 0.63   time = 17587   
#>  22/100   risk = 0.61  oob_risk = 0.63   time = 19412   
#>  24/100   risk = 0.61  oob_risk = 0.62   time = 21232   
#>  26/100   risk = 0.6  oob_risk = 0.62   time = 23050   
#>  28/100   risk = 0.6  oob_risk = 0.62   time = 24912   
#>  30/100   risk = 0.59  oob_risk = 0.61   time = 26969   
#>  32/100   risk = 0.59  oob_risk = 0.61   time = 28846   
#>  34/100   risk = 0.59  oob_risk = 0.61   time = 30707   
#>  36/100   risk = 0.58  oob_risk = 0.61   time = 32581   
#>  38/100   risk = 0.58  oob_risk = 0.61   time = 34444   
#>  40/100   risk = 0.57  oob_risk = 0.6   time = 36301   
#>  42/100   risk = 0.57  oob_risk = 0.6   time = 38191   
#>  44/100   risk = 0.57  oob_risk = 0.6   time = 40049   
#>  46/100   risk = 0.56  oob_risk = 0.6   time = 42420   
#>  48/100   risk = 0.56  oob_risk = 0.6   time = 44281   
#>  50/100   risk = 0.56  oob_risk = 0.59   time = 46193   
#>  52/100   risk = 0.55  oob_risk = 0.59   time = 48061   
#>  54/100   risk = 0.55  oob_risk = 0.59   time = 49939   
#>  56/100   risk = 0.55  oob_risk = 0.59   time = 51786   
#>  58/100   risk = 0.54  oob_risk = 0.59   time = 53649   
#>  60/100   risk = 0.54  oob_risk = 0.59   time = 55540   
#>  62/100   risk = 0.54  oob_risk = 0.58   time = 57405   
#>  64/100   risk = 0.54  oob_risk = 0.58   time = 59272   
#>  66/100   risk = 0.53  oob_risk = 0.58   time = 61121   
#>  68/100   risk = 0.53  oob_risk = 0.58   time = 62984   
#>  70/100   risk = 0.53  oob_risk = 0.58   time = 64842   
#>  72/100   risk = 0.52  oob_risk = 0.58   time = 66717   
#>  74/100   risk = 0.52  oob_risk = 0.57   time = 68582   
#>  76/100   risk = 0.52  oob_risk = 0.57   time = 70456   
#>  78/100   risk = 0.52  oob_risk = 0.57   time = 72308   
#>  80/100   risk = 0.51  oob_risk = 0.57   time = 74179   
#>  82/100   risk = 0.51  oob_risk = 0.57   time = 76027   
#>  84/100   risk = 0.51  oob_risk = 0.57   time = 77882   
#>  86/100   risk = 0.51  oob_risk = 0.57   time = 79713   
#>  88/100   risk = 0.5  oob_risk = 0.56   time = 81639   
#>  90/100   risk = 0.5  oob_risk = 0.56   time = 83459   
#>  92/100   risk = 0.5  oob_risk = 0.56   time = 85295   
#>  94/100   risk = 0.5  oob_risk = 0.56   time = 87130   
#>  96/100   risk = 0.49  oob_risk = 0.56   time = 88958   
#>  98/100   risk = 0.49  oob_risk = 0.56   time = 90788   
#> 100/100   risk = 0.49  oob_risk = 0.56   time = 92609   
#> 
#> 
#> Train 100 iterations in 0 Seconds.
#> Final risk based on the train set: 0.49

file = "cboost.json"
cboost$saveToJson(file, rm_data = TRUE)

cboost_without_data = Compboost$new(file = file)

# The data field now just contains a dummy:
cboost_without_data$data
#>   V1 V10 V11 V12 V13 V14 V15 V16 V17 V18 V19 V2 V20 V21 V22 V23 V24 V25 V26 V27
#> 1  0   0   0   0   0   0   0   0   0   0   0  0   0   0   0   0   0   0   0   0
#>   V28 V29 V3 V30 V31 V32 V33 V34 V35 V36 V37 V38 V39 V4 V40 V41 V42 V43 V44 V45
#> 1   0   0  0   0   0   0   0   0   0   0   0   0   0  0   0   0   0   0   0   0
#>   V46 V47 V48 V49 V5 V50 V51 V52 V53 V54 V55 V56 V57 V58 V59 V6 V60 V7 V8 V9
#> 1   0   0   0   0  0   0   0   0   0   0   0   0   0   0   0  0   0  0  0  0

Note: It is not possible to use any functionality that requires the training data when storing and loading the object without data. For example, cboost$predict() now throws an error:

cboost_without_data$predict()
#> Error in eval(expr, envir, enclos): Production mode is on, this does not allow prediction on training data and hence also blocks the continuation of the training. This is most likely because the training data was removed to either store memory or due to privacy reasons.

Functionality of a data free model

The most important functions are still usable:

Extracting feature importance.

vip = cboost_without_data$calculateFeatureImportance()

Predict on new data

ndat = dat[1:10, ]
cboost_without_data$predict(ndat)
#>              [,1]
#>  [1,] -0.17232799
#>  [2,]  0.18369305
#>  [3,] -0.30361262
#>  [4,] -0.02937112
#>  [5,]  0.64950025
#>  [6,]  0.38152391
#>  [7,]  0.10316929
#>  [8,]  0.98339434
#>  [9,]  0.07043071
#> [10,] -0.53720517

Visualize partial feature effects.

library(patchwork)

# Use most important base learner:
bln = vip$baselearner[1]
plotBaselearner(cboost_without_data, bln) +
plotPEUni(cboost_without_data, strsplit(bln, "_")[[1]][1])

Get logger data

head(cboost_without_data$getLoggerData())
#>   _iterations  oob_risk time baselearner train_risk
#> 1           0        NA   NA   intercept  0.6795747
#> 2           1 0.6771422    0  V11_spline  0.6757390
#> 3           2 0.6742815 1007  V11_spline  0.6719943
#> 4           3 0.6714987 1931  V11_spline  0.6683382
#> 5           4 0.6687920 2819  V11_spline  0.6647688
#> 6           5 0.6661591 3719  V11_spline  0.6612840

Setting the model to a previous iteration.

table(cboost_without_data$getSelectedBaselearner())
#> 
#> V10_spline V11_spline V12_spline V21_spline V36_spline V48_spline V49_spline 
#>         13         20         24         21         11          3          8
cboost_without_data$predict(ndat)
#>              [,1]
#>  [1,] -0.17232799
#>  [2,]  0.18369305
#>  [3,] -0.30361262
#>  [4,] -0.02937112
#>  [5,]  0.64950025
#>  [6,]  0.38152391
#>  [7,]  0.10316929
#>  [8,]  0.98339434
#>  [9,]  0.07043071
#> [10,] -0.53720517

# State after 50 iteration:
cboost_without_data$train(50)
table(cboost_without_data$getSelectedBaselearner())
#> 
#> V10_spline V11_spline V12_spline V21_spline V36_spline 
#>          6         19         15          9          1
cboost_without_data$predict(ndat)
#>              [,1]
#>  [1,] -0.02334412
#>  [2,]  0.28855992
#>  [3,] -0.05161347
#>  [4,] -0.06229294
#>  [5,]  0.60796609
#>  [6,]  0.51006267
#>  [7,]  0.14361074
#>  [8,]  0.79343112
#>  [9,] -0.03531843
#> [10,] -0.29024662

Advantages

Size

The size of the model and JSON file is much smaller when the data is not stored.

file_full = "cboost_full.json"
cboost$saveToJson(file_full)

file.info(file)$size / 1024^2
#> [1] 1.301947
file.info(file_full)$size / 1024^2
#> [1] 3.343959

Loading

Loading a model is much faster (maybe not that striking for smaller models):

system.time(Compboost$new(file = file))
#>    user  system elapsed 
#>   0.101   0.000   0.101
system.time(Compboost$new(file = file_full))
#>    user  system elapsed 
#>   0.145   0.004   0.149

Privacy

Raw data is not shared unintentionally with third parties. This is especially striking for domains that works with sensitive data.