Mitigating sampling bias in risk-based active learning via an EM algorithm