Scalable Reasoning Infrastructure for Large Scale Knowledge Bases

Hima P. Karanam, Sumit Neelam, Udit Sharma, Sumit Bhatia, Srikanta Bedathur, L. Venkata Subramaniam, Maria Chang, Achille Fokoue-Nkoutche, Spyros Kotoulas, Bassem Makni, Mariano Rodriguez-Muro, Ryan Musa, Michael WitbrockISWC (P&D/Industry/BlueSky) 2018 · OpenReview

Abstract

As Artificial Intelligence moves toward general intelligence, the demand for structured knowledge representation is increasing. However, traditional reasoning systems often fail to scale to Knowledge Graphs (KGs) containing billions of facts. This paper introduces a modular, scalable infrastructure that leverages property graphs and optimized query translation to perform complex reasoning over massive datasets. By integrating symbolic reasoners with a distributed graph backend, the system achieves high performance for selective queries while providing a framework for future neural-symbolic integration.

Introduction

The "Reasoning Gap" in modern AI stems from three primary challenges: scalability, expressivity, and multi-modality. Traditional OWL2 reasoners are typically bound to single-machine memory, limiting their utility in big data environments. Furthermore, real-world data often requires handling spatial/temporal context and uncertain information that standard triple stores struggle to manage efficiently. To address these issues, researchers at IBM have developed a toolkit designed for industrial-scale reasoning that maintains the flexibility to handle complex, multi-modal data.

System Architecture

The core of the proposed infrastructure is a transition from traditional triple stores to a Property Graph model, specifically utilizing JanusGraph. This choice allows for the storage of rich metadata—such as context, confidence scores, and evidence—directly as node and edge properties.

Figure 1: Overview of the scalable reasoning architecture, highlighting the ingestion pipeline and query processing layers.

As shown in Figure 1, the architecture consists of several key components:

  1. Ingestion Pipeline: An extension of the RDF4J framework enables the direct ingestion of OWL, RDF, CSV, and text data into the property graph model.
  2. Query Engine & Translation: To bridge the gap between semantic web standards and graph traversals, a translation layer converts SPARQL queries into optimized Gremlin. This layer employs a "Match" algorithm to reorder query patterns based on selectivity, leveraging Gremlin’s internal optimizer.
  3. Hybrid Reasoning: The system integrates symbolic reasoners (e.g., Racer) for TBox (schema) inference and query rewriting. The design is modular, allowing for the future inclusion of neural and IR-based reasoners as fallback mechanisms when symbolic logic fails to find an answer.

Evaluation and Results

The system was evaluated using the LUBM (Lehigh University Benchmark) 100U dataset, representing 100 universities. Performance was compared against Virtuoso, a high-performance triple store, across a suite of standard SPARQL queries.

Figure 2: Performance comparison (in ms) between Virtuoso and JanusGraph (Manual and Auto-translated) across LUBM queries.

The results in Figure 2 reveal a distinct performance split based on query type:

  • Selective Queries: For queries with high filtering capability (e.g., Q1, Q3, Q10), the JanusGraph architecture was orders of magnitude faster than Virtuoso. For instance, Q3 executed in 1.5ms on JanusGraph compared to 55ms on Virtuoso.
  • Large-Scale Scans: Conversely, Virtuoso dominated in queries requiring extensive node scans (e.g., Q2, Q6, Q9). Q2 took approximately 1,006ms on Virtuoso but over 220,000ms on JanusGraph. This highlights the "locality challenge" in distributed graphs, where jumping across disks for non-selective data retrieval introduces significant overhead.

Discussion and Future Work

The evaluation demonstrates that while property graphs are superior for deep relationship exploration and join-intensive tasks, they currently struggle with low-selectivity scans. A major insight from this work is the impact of "physicality" in data storage; in a distributed graph, vertices are spread across disks, making massive scans inefficient compared to traditional relational-style stores.

Future work will focus on improving "locality of reference" through schema transformation techniques to ensure related data is physically stored closer together. Additionally, the team aims to implement ABox (assertion) summarization to handle even larger datasets and integrate neural models to provide probabilistic "best guesses" based on embeddings when symbolic logic is insufficient.

Conclusion

This research provides a viable path toward scaling complex reasoning to the level required for modern AI applications. By combining the strengths of property graphs with optimized SPARQL-to-Gremlin translation, the infrastructure offers a scalable alternative to traditional reasoning engines, particularly for deep, relationship-heavy queries.

Summary updated

BibTeX
@inproceedings{karanam2018scalable,
  author = {Hima P. Karanam and Sumit Neelam and Udit Sharma and Sumit Bhatia and Srikanta Bedathur and L. Venkata Subramaniam and Maria Chang and Achille Fokoue-Nkoutche and Spyros Kotoulas and Bassem Makni and Mariano Rodriguez-Muro and Ryan Musa and Michael Witbrock},
  title = {{Scalable Reasoning Infrastructure for Large Scale Knowledge Bases}},
  booktitle = {ISWC (P\&D/Industry/BlueSky)},
  year = {2018},
  url = {https://openreview.net/forum?id=j-T7400IELi},
}
Figure