Table 2. A socio-technical approach for bias mitigation across the CRISP-DM development cycle.

	Business understanding	Data understanding	Data preparation	Modelling	Evaluation	Deployment
Measurement bias	Team diversity, exchange with domain expert	Proxy estimation	Rapid prototyping
Social bias			Learning fair representation, rapid prototyping, reweighting, optimized preprocessing, data massaging, disparate impact remover	Adversarial debiasing, multiple models, latent variable model, model interpretability equalized odds, prejudice remover
Sampling bias				Resampling		Randomness
Representation bias	Team diversity	Data plotting, exchange with domain experts	Reweighing, data augmentation	Model interpretability
Negative bias			Cross dataset generalization	Bag of words
Label bias		Exchange with domain experts	Data massaging
Sample selection bias			Reweighing
Confounding bias						Randomness
Design Bias			Rapid prototyping	Exchange with domain experts, resampling, model interpretability, multitask learning
Sample treatment bias				Resampling	Data augmentation
Human evaluation bias				Resampling	Representative benchmark subgroup validity, data augmentation
Test dataset bias			Data augmentation
Deployment bias	Team diversity, consequences in context		Rapid prototyping			Monitoring plan, human supervision
Feedback bias						Human supervision, randomness