MiniClean: A Single-Machine System for Cleaning Big Graphs

We demonstrate MiniClean, a single-machine system for cleaning large-scale graphs. MiniClean adopts a rule-based approach that unifies logic reasoning and machine learning, and supports rule discovery, error detection and error correction. To cope with repeated graph pattern matching and large intermediate results, MiniClean proposes (1) a pipelined workflow that coordinates CPU, GPU and I/O operations, (2) memory footprint reduction by bundled processing and data compression, and (3) a multi-mode parallel model that combines SIMD, pipelined, and independent parallelism to maximize CPU--GPU synergy. We demonstrate how MiniClean cleans real-life billion-scale graphs and outperforms a SOTA multi-machine system with a 32-node cluster by at least 8.09X.