In multi-instance learning, instances are organized into bags, anda bag is labeled positive if it contains at least one positive instance, and neg-ative otherwise; the labels of the individual instances are not given. The taskis to learn a classifier from this limited information. While the original taskdescription involved learning an instance classifier, in the literature the taskis often interpreted as learning a bag classifier. Depending on which of thesetwo interpretations is used, it is more natural to evaluate classifiers accordingto how well they predict, respectively, instance labels or bag labels. In theliterature, however, the two interpretations are often mixed, or the intendedinterpretation is left implicit. In this paper, we investigate the difference be-tween bag-level and instance-level accuracy, both analytically and empirically.We show that there is a substantial difference between these two, and bet-ter performance on one does not necessarily imply better performance on theother. It is therefore useful to clearly distinguish them, and always use theevaluation criterion most relevant for the task at hand.
展开▼