The concept of algorithmic fairness is intricate and multifaceted, concerning the equitable treatment of individuals or groups by algorithms in data processing and decision-making. This issue not only impacts trust in algorithms but also influences societal acceptance and reliance on technology. With the growing utilisation of big data and machine learning in predictive algorithms, crime prediction software has become integral to modern policing. However, data bias within these algorithms can result in unfair outcomes, affecting policing decisions, undermining the perceived fairness of algorithmic choices and exacerbating social inequality. Current crime prediction software exhibits three primary types of bias: design bias, training bias and interaction bias. These biases are ingrained in algorithms through predictive models and materialise as feedback loops in practice. Mitigating the adverse effects of data biases can be achieved by enhancing algorithmic transparency and interpretability, enhancing dataset diversity and fairness, and optimising feedback mechanisms to guide the evaluation of future algorithmic fairness practices.
Fan et al. (Thu,) studied this question.