Bayesian Sampler: Fairness in Sampling

We reinterpret the concept of Bayesian balance from Yang Liu [26] to find connection between fairness of algorithms and class imbalance in latent clusters of training data. We argue that the degree of class imbalance in the latent clusters of some training data is manifested in the lack of fairness of an algorithm trained on that data. A novel algorithm is proposed in this paper which decides an optimal policy to draw a balanced data sample from clustered raw data with class imbalance. The proposed Bayesian network model trades off accuracy with redefined Bayesian balance to find the optimal policy. We claim that a novel application of this sampling technique would be sampling a fair training set for recommender systems. To illustrate, we present experimental results on two real world recommender systems raw data sets and one synthetic data-set.