Data preprocessing to mitigate bias: A maximum entropy based approach