In data_analysis_2.ipynb I can find the approach to improve the subgroup discovery by upsampling some of the attributes in the data:
Deal with imbalanced data
"high_calorie": "medium_calorie": "low_calorie" ratio is around 2.5:1:1, the data is quite imblanced with regard to the calorie level, we want to try to resample the dataset to help improve the data quality. We want to see if we can improve the quality of subgroup discovery
- upsampling: upsample "medium_calorie" and "low_calorie" samples to make them equal to the number of "high_calorie"
My question is: is this (really) a valid approach to discover new subgroups? I would expect that changing the structure of my data will deliver artificial subgroups that do not exist in reality.
If this is really a valid approach, a reference to a paper / blog etc. would be very helpful, too.