MICCLLR: Multiple-Instance Learning Using Class Conditional Log Likelihood Ratio

Multiple-instance learning (MIL) is a generalization of the supervised learning problem where each training observation is a labeled bag of unlabeled instances. Several supervised learning algorithms have been successfully adapted for the multiple-instance learning settings. We explore the adaptation of the Naive Bayes (NB) classifier and the utilization of its sufficient statistics for developing novel multiple-instance learning methods. Specifically, we introduce MICCLLR (multiple-instance class conditional log likelihood ratio), a method for mapping each bag of instances as a single meta-instance using class conditional log likelihood ratio statistics such that any supervised base classifier can be applied to the meta-data. The results of our experiments with MICCLLR using different base classifiers suggest that no single base classifier consistently outperforms other base classifiers on all data sets. We show that a substantial improvement in performance is obtained using an ...

• Type: Conference
• Year: 2009
• Where: DIS
• Authors: Yasser El-Manzalawy, Vasant Honavar
• Keywords: Base Classifiers, DIS 2009, Multiple-Instance Learning, Supervised Learning, Theoretical Computer Science