Instance-level accuracy versus bag-level accuracy in multi-instance learning

Multi-instance learning is a learning task where limited information about instance labels is given: instances are organized into bags, and a bag is labeled positive if it contains at least one positive instance, and negative otherwise; the labels of the individual instances are not given. The task is to learn a classifier from this limited information. While the original task description involved the learning of an instance classifier, in the literature the task is often interpreted as learning a bag classifier. Similarly, classifiers (and the corresponding learners) can be evaluated based on the accuracy with which they classify instances, or bags. In the literature, the two different settings (instance versus bag classification) are often intermingled. In this paper, we investigate more closely the difference between bag-level and instance-level accuracy, both analytically and empirically. We show that there is a large difference in practice between these two, and results for bag-level accuracy do not necessarily hold for instance-level accuracy (and vice versa). It is therefore useful to clearly distinguish these two, and use the most suitable one for the task at hand.

Publication year:2011

Accessibility:Open