Constant columns labeled as continuous

[amueller]

The training set contains several constant columns labeled as continuous:

train_merged[(train_merged.y_act == 0) & (train_merged.num_of_dist_val == 1)]

     Record_id             Attribute_name  y_act  total_vals  num_nans  \
142         51                      count      0       29421         0   
587        101  M1_CURRENT_PROGRAM_NUMBER      0         605         0   
627        102  M1_CURRENT_PROGRAM_NUMBER      0         740         0   
640        102           S1_SystemInertia      0         740         0   

     %_nans  num_of_dist_val  %_dist_val  mean  std_dev  min_val  max_val  \
142     0.0                1    0.003399   1.0      0.0      1.0      1.0   
587     0.0                1    0.165289   1.0      0.0      1.0      1.0   
627     0.0                1    0.135135   1.0      0.0      1.0      1.0   
640     0.0                1    0.135135  12.0      0.0     12.0     12.0   

    sample_1 sample_2 sample_3 sample_4 sample_5  \
142        1        1        1        1        1   
587        1        1        1        1        1   
627        1        1        1        1        1   
640       12       12       12       12       12   

                                name  \
142  aac_shelter_cat_outcome_eng.csv   
587                experiment_08.csv   
627                experiment_09.csv   
640                experiment_09.csv   

                                                  link  
142  https://www.kaggle.com/aaronschlegel/austin-an...  
587  https://www.kaggle.com/shasun/tool-wear-detect...  
627  https://www.kaggle.com/shasun/tool-wear-detect...  
640  https://www.kaggle.com/shasun/tool-wear-detect...  

I think these should be labeled as not-generalizable.

[amueller]

Actually, they are labeled as all kinds of things, mostly lists:

train_merged[(train_merged.y_act != 6) & (train_merged.num_of_dist_val == 1)].y_act.value_counts()

7    327
2     12
4      5
0      4
3      3
1      1

[pvn25]

Thank you for pointing this out. I found 25 examples overall that should have been Not-Generalizable in our labeled data. I will update the dataset along with the benchmark with the changes.

So, 6 is decoded as List and 7 as Not-Generalizable. I would like to apologize for this encoding, it is a bit confusing. I will change them to actual types.

[amueller]

Oh sorry for the 6/7 confusion, I guess it wasn't as bad as I made it look :) Thanks for checking up on it!