Avoiding False Positive in Multi-Instance Learning