Scalability: Traditional approaches to machine learning depend on samples being statistically independent. Because the samples are statistically independent, data scientists can isolate the individual contribution of each subset, and thereby optimize training. In contrast, in graph data structures the nodes are interconnected, so they are not statistically independent. The interrelatedness between nodes in graph data structure creates challenges in sampling, because the subgraphs sampled need to maintain a representative structure, but the interconnectedness can introduce bias into the training sample that distorts representativeness (e.g., nodes and edges that appear more often than others in the training set).xxi