Rebalance columns

With MOSTLY AI, you can generate synthetic datasets that rebalance column distributions from your original data. You can use rebalancing for fairer distributions and bias mitigation in your synthetic datasets..

Rebalancing does not just impact the single variable you select to rebalance. All other variables in the dataset are also impacted based on the correlations they have with that variable. For example, you can use rebalancing to create a large number of relevant business scenarios out of the few that are present in your original data. Or you can use it to simulate "what-if" scenarios based on your existing, historical data, or upsample minority classes to help downstream machine learning algorithms pick up their patterns.

Prerequisites

To rebalance columns in generated synthetic datasets, the model you use for rebalancing must have flexible generation enabled.

Specify columns and values to rebalance

When you configure a new synthetic dataset, you can specify the columns you want to rebalance.

You can examine the CDNOW dataset where the distributions in the gender column of the customers table is 58% males and 42% females.

The examples below show how to rebalance the customers column and achieve a 50% - 50% distribution of males and females in the generated synthetic data.

# train generator on imbalanced data
g = mostly.train(config={
  "name": "US census",
  "tables": [
    {
      "name": "census",
      "data": "https://mostly.ai/docs/datasets/us-census-income.csv"
    }
  ]
})
 
# generate 2k balanced synthetic samples
sd = mostly.generate(g, config={
   "name": "US census rebalanced",
   "tables": [
      {
         "name": "census",
         "configuration": {
            "rebalancing": {
               "column": "income",
               "probabilities": {
                  ">50K": 0.5
               }
            }
         }
      }
   ]
})
df_synthetic = sd.data()
df_synthetic.income.value_counts()