There's a proof of concept naive Bayesian approach here: https://github.com/MarcT0K/Fediverse-Spam-Filtering