DALEX on custom stacked models

[leungi]

Pseudo-reprex below to illustrate workflow.

There are 2 stages of stacking, and below is abbreviated to final stage.

# input data for prediction; this data are themselves a result of stacked model
df <- tibble::tribble(
  ~x1, ~x2, ~true,           ~pred1,           ~pred2,           ~pred3,           ~pred4,           ~pred5,
  "0016",   1, 11255, 9782.06546666667, 8226.73783726366, 8423.53411898339, 7663.85714285714, 7778.32234611454,
  "0016",   2, 10155, 9917.16225000001,  7390.2726470072, 7548.50621212894, 6011.57142857143,  7020.0197927677,
  "0016",   3,  9905, 8365.66048333333, 4748.35733132711, 4897.40398331136, 5625.14285714286, 5197.59820269678,
  "0026",   1,  9569, 10542.7790333333, 12448.8281473898, 12982.2853847065, 9529.42857142857, 9913.60100542533,
  "0026",   2, 15004,      12332.88455, 13118.3179554928, 13490.4519001908, 9449.14285714286, 9782.48187764126,
  "0027",   1,  4623, 6228.92556666668, 7901.02224985066,  8072.3059097473, 7663.85714285714,  7564.7019858157,
  "0027",   2,  3666, 3902.33416666666, 5351.58779239503, 5501.55032427708, 5757.85714285714, 5791.90612060224,
  "0027",   3,  2046, 3730.91108333333, 5405.90164588071, 5431.22100425988,             5700, 5574.85787520228,
  "0345",   1,  7848, 7911.66811666667, 7332.14726332333, 7535.03388134704, 8428.85714285714, 7504.20919309283,
  "0345",   2,  5594,        6249.8431, 5302.09068924222, 5602.24650648537,             6253, 5936.17306199591,
  "0348",   1,  6118,        5888.9112,  6782.1549012783, 6983.85792156352, 7145.28571428571, 6996.64665890851,
  "0348",   2,  4115,        4655.3621, 4061.92478416692,  4339.3944039624, 5379.71428571429, 5201.36079952954,
  "0348",   3,  3792, 4703.56786666666, 4862.77758785772, 4886.36623749198, 5413.85714285714,  5316.2047603152,
  "1000",   1,  9982,        8894.2428, 8950.05680053561, 8724.27457157357, 7643.14285714286, 8427.52273508174,
  "1000",   2,  4218, 5103.73553333333, 6755.30317981863, 6492.15505744351,             7836, 6900.52725335413,
  "1022",   1,  9021, 8966.84941666667, 8921.14926298024, 8514.45660876879, 8590.57142857143, 8566.07119574923,
  "1022",   2, 11692, 10205.8180333333, 8895.88440879051, 8417.59814231434, 8185.85714285714, 8225.60579235643,
  "1022",   3,  9420, 9664.82173333334, 9422.99681882565, 8835.71873031759, 7853.57142857143, 8126.76078652109,
  "1022",   4,  6850, 7419.07043333333, 8995.48869657391, 8194.63910112673, 7604.14285714286, 7815.14405713875,
  "1022",   5,  6850, 7419.07043333333,  8817.8438463534, 7883.22080414475, 6846.14285714286, 7515.84608489043
)

# model list for stacking
md <- list(rf,
           pca,
           svm,
           enet)

model_pred_stack <- function(df, md) {

  # iterate over list of models in md, and average prediction
  temp <- 0
  for (i in 1:length(md)) {
    temp <- temp + predict(md[[i]], df)
  }
  temp <- temp / length(md)
  return(temp)
}

model_pred <- model_pred_stack(df = df, md)

# with DALEX, have to loop over list of models one by one, which doesn't reflect intention of stacking; otherwise, modelDown will complain
explain_stacked <- explain(
  md[[1]],
  data = df,
  y = df$true,
  label = "stacked"
)

modelDown(explain_stacked,
          device = "svg",
          output_folder = "output_data/modelDown_stacked")

# passing list of models; modelDown fails to generate diagnostics, aside from data description
explain_stacked <- explain(
  md,
  data = df,
  y = df$true,
  label = "stacked"
)

modelDown(explain_stacked,
          device = "svg",
          output_folder = "output_data/modelDown_stacked")

[kromash]

Thank you for provided example. I will look into it.

[leungi]

Appreciate the prompt response; your team's work (starting from DALEX) is a game changer 👍