A meta-learning system for multi-instance classification

Meta-learning refers to the use of machine learning methods to analyze the behavior of machine learning methods on different types of datasets. Until now, meta-learning has mostly focused on the standard classification setting. In this ongoing work, we apply it to multiinstance classification, an alternative classification setting in which bags of instances, rather than individual instances, are labeled. We define a number of data set properties that are specific to the multi-instance setting, and extend the concept of landmarkers to the multi-instance setting. Experimental results show that multi-instance classifiers are very sensitive to the context in which they are used, and that the meta-learning approach can indeed yield useful insights in this respect.