Plant-herbivore interaction networks provide information about community organization. Two methods are currently used to document pairwise interactions among plants and insect herbivores. One is the traditional method that collects plant-herbivore interaction data by field observation of insect occurrence on host plants. The other is the increasing application of newly developed molecular techniques based on DNA barcodes to the analysis of gut contents. The second method is more appealing because it documents realized interactions. To construct complete networks, each technique of network construction is urgent to be assessed. We addressed this question by comparing the effectiveness and reliability of the two methods in constructing plant-Lepidoptera larval network in a 50 ha subtropical forest in China. Our results showed that the accuracy of diet identification by observation method increased with the number of observed insect occurrences on food plants. In contrast, the molecular method using three plant DNA markers were able to identify food residues for 35.6% larvae and correctly resolved 77.3% plant (diet) species. Network analysis showed molecular networks had three-fold more unique host plant species but fewer links than the traditional networks had. The molecular method detected plants that were not sampled by the traditional method, e.g., bamboos, bryophytes and lianas in the diets of insect herbivores. The two networks also possessed significantly different structural properties. Our study indicates the traditional observation of co-occurrence is inadequate, while molecular method can provide higher species resolution of ecological interactions.