Evaluation metrics play an important role in classifiers.Popular evaluation metrics used in multi-label learning include Hamming loss,One-error,Coverage,Ranking loss and Average precision.While the classification results are obtained from multi-label classifier,the val-ues of evaluation metrics will be derived later,usually the evaluation metrics are assessed in the way of checking afterwards.However this sometimes cannot find the problem of the variation in values of evaluation metrics timely and effectively,meanwhile it is necessary to mark the test samples when estimating the values of evaluation metrics.To solve this problem,this paper put forward two methods of estimating the eval-uation metrics based on the difference in sample sets distribution and on the difference between instances in sample sets respectively.After an-alysing the characteristics of above two methods,we propose the third estimating method for evaluation metrics.Experiments show that the pro-posed three methods all have good effects.They can be used in transfer learning and others.%评价标准是分类器的重要指标。对于多标签学习，常用的评价标准有 Hamming Loss、One-error、Coverage、Ranking loss 和Average precision。多标签分类器给出分类结果的同时并未给出评价标准值，通常采用事后验算的方法评估评价标准。这样往往不能及时有效地发现评价标准值变化之类的问题，同时评估评价标准值需对测试样本进行标记。针对这一问题，分别从样本分布差异和样本实例间差异提出两种评价标准预估方法。分析上述两种方法的特点，提出第三种评价标准预估方法。实验表明，这三种评价标准预估方法具有良好效果，可用于迁移学习等。
展开▼
