Privacy requirements often restrict organizations from distributing sensitive data for things like running a public hackaton, PoC evaluation with external organizations, and even product development at different units within the originating organization.

Synthetic data are simulated data designed to mimic the properties of a real-world dataset. Synthetic data are often used as externally-available alternatives to sensitive data. For example,

Generating synthetic data with the Redpoll Reformer

If you have a Reformer server running, generating a csv of synthetic data four lines of python code:

import redpoll
rp = redpoll.Client(client_address)
# Generate 1000 rows of synthetic data
df = rp.simulate(redpoll.AllColumns, n=1000)
df.to_csv("my-synthetic-data.csv")

Synthetic data with balanced classes

Class imbalance is a big issue in real-world datasets.

import redpoll
import pandas as pd
rp = redpoll.Client(client_address)
# Generate 1000 rows of synthetic binary classification data with 
# 500 examples of each class
df0 = rp.simulate(redpoll.AllColumns, given={'class': 0}, n=500)
df1 = rp.simulate(redpoll.AllColumns, given={'class': 1}, n=500)
pd.concat([df0, df1]).to_csv("my-balanced-synthetic-data.csv")