This paper investigates methods to modify the naive Bayes classifier to ensure that its predictions are independent of a sensitive attribute, such as gender or race, which may be associated with discrimination. The authors present three approaches: (i) modifying the probability of the positive class given the sensitive attribute, (ii) training separate models for each sensitive attribute value and balancing them, and (iii) adding a latent variable to the Bayesian model to represent unbiased labels and optimizing the model parameters using expectation maximization. The paper motivates the problem by discussing the red-lining effect, where correlated attributes are used to indirectly discriminate. Experiments on artificial and real-life datasets show that the two-models approach outperforms the latent variable method in terms of accuracy and discrimination. The paper also discusses related work and future directions, including extending the notion of discrimination to conditional discrimination and exploring other graphical models.This paper investigates methods to modify the naive Bayes classifier to ensure that its predictions are independent of a sensitive attribute, such as gender or race, which may be associated with discrimination. The authors present three approaches: (i) modifying the probability of the positive class given the sensitive attribute, (ii) training separate models for each sensitive attribute value and balancing them, and (iii) adding a latent variable to the Bayesian model to represent unbiased labels and optimizing the model parameters using expectation maximization. The paper motivates the problem by discussing the red-lining effect, where correlated attributes are used to indirectly discriminate. Experiments on artificial and real-life datasets show that the two-models approach outperforms the latent variable method in terms of accuracy and discrimination. The paper also discusses related work and future directions, including extending the notion of discrimination to conditional discrimination and exploring other graphical models.