Drawer Graph Neural Network for Multi-Instance Learning

Multi-instance Learning (MIL) aims to classify data bags, while the labels of individual instances within each bag are unknown. Traditional MIL methods take the bag as the fundamental unit of analysis, classifying each bag exclusively based on the features of its constituent instances. However, when different combinations of instances form bags with distinct labels, these methods face a critical challenge: instances are isolated by bag boundaries, which prevents the model from capturing the subtle discrepancies between instance combinations across bags and thereby degrades the accuracy of bag label prediction. To tackle this issue, we design a MIL framework that takes instances as the basic analysis unit and breaks the constraints imposed by bag boundaries, explicitly extending instance correlation analysis from the intra-bag scope to the holistic instance space across bags. Under this paradigm, we propose the drawer graph neural network for MIL (DragnMIL), which formalizes the holistic instance space as a drawer graph, where bags serve as compartments that partition the space and instances are distributed among them. Bag classification is achieved by learning instance features and correlations over the drawer graph, so DragnMIL models both intra-bag instance relations and cross-bag instance correlations, enabling finer-grained discrimination of instance combinations and stronger bag classification performance. Experimental results validate that DragnMIL is effective compared with state-of-the-art methods on multiple MIL datasets.