<p>In Multi-Instance Learning (MIL), a sample is viewed as a bag consisting of instances. Traditional approaches often rely on the assumption that the instances in a bag are independently and identically distributed (i.i.d.), thereby neglecting the potential dependencies among instances. In recent years, increasing attention has been paid to the non-i.i.d. setting, leading to the development of a new variety of methods, i.e., graph-based MIL. This paper provides a systematic review of two representative categories of graph-based MIL methods under the non-i.i.d. assumption: the first category centers around graph kernels, which construct graph structures to capture instance-level dependencies and design kernel functions to measure similarity between bags; the second category leverages Graph Neural Networks (GNNs), mapping each bag into a fixed-dimensional embedding through graph construction and aggregation mechanisms. We analyze the modeling ideas and core techniques of both groups of methods and categorize existing approaches accordingly. In addition, we summarize real-world applications of graph-based MIL, highlighting why non-i.i.d. structures naturally arise and how graph modeling benefits these tasks. Finally, we make a discussion on the current challenges in this line of research and outline potential future directions.</p>