Given a large enough volume of data and precise, meaningful categories, training a statistical model to solve a classification problem is straightforward and has become a standard application of machine learning (ML). If the categories are not precise, but rather fuzzy, as in the case of scientific disciplines, the systematic failures of ML classification can be informative about properties of the underlying categories. Here we classify a large volume of academic publications using only the abstract as information. From the publications that are classified differently by journal categories and ML categories (i.e., misclassified publications, when using the journal assignment as ground truth) we construct a network among disciplines. Analysis of these misclassifications provides insight in two topics at the core of the science of science: (1) Mapping out the interplay of disciplines. We show that this misclassification network is informative about the interplay of academic disciplines and it is similar to, but distinct from, a citation-based map of science, where nodes are scientific disciplines and an edge indicates a strong co-citation count between publications in these disciplines. (2) Analyzing the success of interdisciplinarity. By evaluating the citation patterns of publications, we show that misclassification can be linked to interdisciplinarity and, furthermore, that misclassified articles have different citation frequencies than correctly classified articles: In the highest 10 percent of journals in each discipline, these misclassified articles are on average cited more frequently, while in the rest of the journals they are cited less frequently.
