Causal Structure Learning of Bias for Fair Affect Recognition