International Conference on Information Systems (ICIS 2021: Building Sustainability and Resilience with IS: A Call for Action)
Artificial intelligence, Data Science -- methods, Algorithms -- Social aspects
Literature on algorithmic bias identifies its source in either biased data or statistical methods, more rarely in the development of algorithmic solutions as a potential factor. Because of the prior unknowability of algorithms, data scientists developing such solutions have to take various design decisions. Drawing from the flow-oriented approach, we study algorithmic unknowability and how data scientists respond to it in 35 public data science Jupyter notebooks containing algorithmic solutions to predict customer churn in a credit card dataset on a data science platform Kaggle.com. We offer a more thorough understanding of the unknowability in algorithmic development that can enable algorithmic bias: resource, problem, dataset, analytical, model, and performance unknowability. We find that in response, data scientists engage in biasenabling interpretation, bias-enabling optionalizing, and bias-enabling experimentation. These findings contribute to literature on algorithmic bias and can help avert bias earlier in practice.
A video of this presentation is available online:
This is the accepted manuscript version of a conference paper. Final version copyrighted © by Association for Information Systems. Authors retain copyright for material published as part of AIS conference proceedings.
Stelmaszak, M. (2021) How To Train Your Algo: Investigating the Enablers of Bias in Algorithmic Development. International Conference on Information Systems, 12-15 December 2021.