Here's where we can start:

• Use of stratified sampling and optimized observation weights.
• Careful analysis of mussing values for systematic bias.
• Standardize processes and create templates for model documentation and deployment.
• Challenger team validation of machine-learning algorithms. 

Alexandra Kory

LinkedIn messenger finishes sentences for you. Have you ever realized that when you refer to a person as a CEO, chances are the algorithm will male pronouns (him, his) more than female ones... why? Because it's learned that! From bias. There in fact are many tangible examples. But they all have in common that the data that were used for training the algorithm came with an inbuilt bias. 

Alexandra Kory