ZeroFolio: Algorithm Selection via Text Embeddings Without Domain Knowledge

ZeroFolio is a new method that replaces the traditional way of creating instance features with pre-trained text embeddings, aimed at algorithm selection. It handles raw text files, uses a pre-trained model to create embeddings, and relies on weighted k-nearest neighbors for algorithm choice. A key advantage is that it doesn’t require any specific knowledge or training for different tasks. When evaluated across 11 ASlib scenarios from 7 domains—like SAT, MaxSAT, QBF, ASP, CSP, MIP, and various graph issues—ZeroFolio outperformed a random forest model based on manually crafted features in 10 out of those 11 scenarios, all while using a single, consistent setup.