Novelty detection under multi-label multi-instance framework

Abstract

Novelty detection plays an important role in machine learning and signal\nprocessing. This paper studies novelty detection in a new setting where the\ndata object is represented as a bag of instances and associated with multiple\nclass labels, referred to as multi-instance multi-label (MIML) learning.\nContrary to the common assumption in MIML that each instance in a bag belongs\nto one of the known classes, in novelty detection, we focus on the scenario\nwhere bags may contain novel-class instances. The goal is to determine, for any\ngiven instance in a new bag, whether it belongs to a known class or a novel\nclass. Detecting novelty in the MIML setting captures many real-world phenomena\nand has many potential applications. For example, in a collection of tagged\nimages, the tag may only cover a subset of objects existing in the images.\nDiscovering an object whose class has not been previously tagged can be useful\nfor the purpose of soliciting a label for the new object class. To address this\nnovel problem, we present a discriminative framework for detecting new class\ninstances. Experiments demonstrate the effectiveness of our proposed method,\nand reveal that the presence of unlabeled novel instances in training bags is\nhelpful to the detection of such instances in testing stage.\n