A crucial step in planet hunting surveys is to select the best candidates for follow-up observations, given limited telescope resources. This is often performed by human 'eyeballing', a time consuming and statistically awkward process. Here, we present a new, fast machine learning technique to separate true planet signals from astrophysical false positives. We use self-organizing maps (SOMs) to study the transit shapes of Kepler and K2 known and candidate planets. We find that SOMs are capable of distinguishing known planets from known false positives with a success rate of 87.0 per cent, using the transit shape alone. Furthermore, they do not require any candidate to be dispositioned prior to use, meaning that they can be used early in a mission's lifetime. A method for classifying candidates using a SOM is developed, and applied to previously unclassified members of the Kepler Objects of Interest (KOI) list as well as candidates from the K2 mission. The method is extremely fast, taking minutes to run the entire KOI list on a typical laptop. We make python code for performing classifications publicly available, using either new SOMs or those created in this work. The SOM technique represents a novel method for ranking planetary candidate lists, and can be used both alone or as part of a larger autovetting code.