The purpose of synthesizing a graph from input data in neural structured learning is to incorporate structured relationships and dependencies among data points into the learning process. By representing the input data as a graph, we can leverage the inherent structure and relationships within the data, which can lead to improved model performance and generalization.