Right now, `co_occurrances` is excruciatingly slow for reasonably-large data. Look at parallelism (and other things?) to speed that up, because we can't be this slow.