CPI with xplainfi with groups

cpi-xplainfi.R

library(xplainfi)
library(mlr3learners)
#> Loading required package: mlr3
# Make sure ranger is installed

# Some data with correlated features
task = sim_dgp_correlated(n = 1000, r = .7)
learner = lrn("regr.ranger", num.trees = 100)
measure = msr("regr.mse")
resampling = rsmp("holdout")

# Create knockoff matrix (Gaussian) from task (uses all data)
knockoff_sampler = KnockoffGaussianSampler$new(task = task)

cfi = CFI$new(
    learner = learner,
    task = task,
    measure = measure,
    resampling = resampling,
    sampler = knockoff_sampler,
    # Could use more repeats for stability but would need to creat emultiple knockoff matrices in sampler
    n_repeats = 1
)

cfi$compute()

# Same as CPI under the hood:
cfi$importance(ci_method = "cpi")
#> Key: <feature>
#>    feature    importance          se   statistic      p.value   conf_lower
#>     <char>         <num>       <num>       <num>        <num>        <num>
#> 1:      x1  4.3122236075 0.369125526 11.68226879 6.020552e-27  3.703367219
#> 2:      x2 -0.0007774337 0.016377543 -0.04746949 5.189162e-01 -0.027791474
#> 3:      x3  1.7174102789 0.117239743 14.64870386 3.970628e-38  1.524028457
#> 4:      x4  0.0027353962 0.002649429  1.03244751 1.513070e-01 -0.001634721
#>    conf_upper
#>         <num>
#> 1:        Inf
#> 2:        Inf
#> 3:        Inf
#> 4:        Inf

# Observation-wise loss differences
cfi$obs_loss()
#>       feature iter_rsmp iter_repeat row_ids loss_baseline    loss_post
#>        <char>     <int>       <int>   <int>         <num>        <num>
#>    1:      x1         1           1       2  1.234542e-06  2.636605010
#>    2:      x1         1           1       4  2.673268e-03 19.545713486
#>    3:      x1         1           1      11  4.150419e-02  0.008964166
#>    4:      x1         1           1      16  7.323025e-04  0.826810619
#>    5:      x1         1           1      29  1.553946e-02  2.216122152
#>   ---                                                                 
#> 1328:      x4         1           1     995  6.389680e-02  0.028051615
#> 1329:      x4         1           1     996  1.995682e-02  0.032137676
#> 1330:      x4         1           1     998  1.184186e+00  1.371750862
#> 1331:      x4         1           1     999  4.278355e-01  0.569745025
#> 1332:      x4         1           1    1000  2.227433e-02  0.010902248
#>       obs_importance
#>                <num>
#>    1:     2.63660377
#>    2:    19.54304022
#>    3:    -0.03254002
#>    4:     0.82607832
#>    5:     2.20058269
#>   ---               
#> 1328:    -0.03584518
#> 1329:     0.01218085
#> 1330:     0.18756487
#> 1331:     0.14190949
#> 1332:    -0.01137208


# Same but with groups:

task$feature_names
#> [1] "x1" "x2" "x3" "x4"

groups = list(
    correlated = c("x1", "x2"),
    noise = c("x3", "x4")
)

cfi = CFI$new(
    learner = learner,
    task = task,
    # Just add the group definition, rest is identical
    groups = groups,
    measure = measure,
    resampling = resampling,
    sampler = knockoff_sampler,
    n_repeats = 1
)

cfi$compute()
cfi$importance(ci_method = "cpi")
#> Key: <feature>
#>       feature importance        se statistic      p.value conf_lower conf_upper
#>        <char>      <num>     <num>     <num>        <num>      <num>      <num>
#> 1: correlated   4.607162 0.3392991  13.57847 5.224602e-34   4.047504        Inf
#> 2:      noise   1.709909 0.1280463  13.35383 3.735263e-33   1.498702        Inf