Cost-Preserving Distributed Approximate Pattern Mining for Scalable RDF Graph Summarization

RDF graph summarization methods based on approximate pattern mining have been shown to produce high-quality summaries. They produce compact summaries while providing explicit control over the trade-off between compression and reconstruction quality. However, existing cost-based approaches remain memory-bound, making them unsuitable for large knowledge graphs that cannot be processed on a single machine. Existing distributed pattern mining techniques improve scalability by partitioning the data and extracting patterns independently, but they cannot preserve the global reconstruction-cost optimization required by costbased approximate pattern mining. In this paper we present a distributed framework for approximate pattern mining that preserves the original optimization objective during parallel execution. Our framework first extracts approximate patterns independently from data partitions and consolidates identical patterns. Since this initial consolidation cannot recover globally optimal patterns spanning multiple partitions, a second cost-preserving merging phase is introduced. To support this phase without rescanning the original dataset, we propose the Missing-Neighbors Index, which records compact local statistics that enable accurate merge-cost evaluation while preserving the semantics of the original reconstruction-cost function. The proposed framework is instantiated for RDF graph summarization using Hadoop MapReduce and evaluated on several real-world RDF datasets. Experimental results demonstrate consistent reductions in execution time while maintaining summary quality nearly identical to the sequential SemSum+ baseline. Although motivated by RDF graph summarization, the proposed framework is applicable to a broader class of cost-based approximate pattern mining methods relying on binary representations and global optimization.