Exploiting Graph Structure to Summarize and Compress Relational Knowledge

AI systems consume volumes of relational knowledge to extract patterns and model the world. We face the challenges of scaling to growing volumes of knowledge, identifying more sophisticated patterns and concepts, and compressing and indexing relational knowledge for future access. This paper describes the domain-general path-specific isomorphic abstraction (ψ-abstraction) approach for encoding relational knowledge as relational graphs, and exploiting paths and cycles in these graphs to guide compression and pattern extraction. ψ-abstraction automatically summarizes relational graphs by (1) identifying cyclic and acyclic paths; (2) generalizing the graph along those paths; (3) explicitly annotating generalizable sequences; (4) encoding qualitative relations over generalized quantity sequences; and (5) optionally pruning generalized knowledge to compress the graph. We demonstrate ψ-abstraction on twelve examples spanning diverse domains: 2D spatial; 3D spatial; temporal (calendar) concepts; abstract feedback cycles; narratives; cell signaling; and anatomy. We evaluate ψ-abstracted knowledge for data compression, pattern identification, and structural similarity of examples within and across categories.