PRISM

Mining rules from relational data is a key problem in AI. The basis of rule mining is mining repeating data patterns, known as structural motifs. Despite the importance of mining ``good” structural motifs, the problem was not well understood. PRISM is the first principled technique for mining structural motifs for learning languages that blend first-order logic with probabilistic models. PRISM is implemented in C++ and can improve over state-of-the-art rule mining techniques, namely LSM and BOOSTR, by up to 6% in terms of accuracy and up to 80% in terms of runtime.

PRISM has been the basis for developing STECTRUM, an even more efficient C++ rule mining technique that scales to databases of million facts. Furthermore, SPECTRUM outperforms on the most well-known benchmarks, such as WN18RR and FB15K-237 state-of-the-art rule mining techniques for entity linking, such as AMIE3, RNNLogic, and NCRL, by up to x100 in terms of runtime, while running exclusively on CPU.

Repository

PRISM

Relevant publications