Towards a Unified Knowledge Graph Data Management System

Baozhu Liu, Xin Wang, Pengkai Liu, Sizhuo LiSEA-Data@VLDB 2021 · OpenReview

Abstract

The Knowledge Graph (KG) ecosystem is currently divided between the RDF model, optimized for web-scale triple-stores, and Property Graphs, designed for complex attribute handling. This fragmentation leads to data redundancy and interoperability challenges. We present a unified KG management system that bridges this gap by utilizing a relational storage scheme and a common semantic framework based on relational algebra. Our system supports both SPARQL and Cypher queries over a single, interoperable storage layer, providing a unified environment for managing, querying, and visualizing diverse graph data.

Introduction

Currently, researchers and engineers must choose between two dominant graph data models: the Resource Description Framework (RDF) and Property Graphs. RDF, a W3C standard often queried via SPARQL, excels at web-scale data integration. Conversely, Property Graphs, popularized by systems like Neo4j, use the Cypher language to manage complex attributes on nodes and edges.

Because these models rely on distinct storage structures and query languages, they are rarely interoperable. Maintaining separate systems for a single project leads to synchronization overhead and a steep learning curve. This paper proposes a system that treats both models as first-class citizens, eliminating "model lock-in."

System Architecture

The proposed system achieves unification through a three-tiered architecture designed for storage agnosticism and query flexibility.

The unified architecture spanning from web interface to relational storage.

1. Unified Storage Manager

Instead of a native graph store, the system "shreds" both RDF and Property Graphs into a relational model. It utilizes a series of Vertex and Edge Tables. To efficiently manage the schema-less nature of graph data, the system employs a characteristic-set-based method to handle untyped entities.

2. Query Processing Coordinator

This layer serves as the system's translator. It decomposes incoming SPARQL or Cypher queries into a unified semantic layer based on relational algebra. By establishing this common mathematical ground, the system can apply standard database optimization techniques to graph queries regardless of their original syntax.

3. Unified Web Interface

The front-end allows users to switch between query languages seamlessly. It provides visualization tools for both query results and the underlying explanations of how those results were derived from the unified storage.

Key Results and Evaluation

The system was evaluated against two state-of-the-art baselines: gStore (an RDF-based engine) and Neo4j (a Property Graph-based engine). The comparison focused on functional parity and query capabilities.

Comparison of features and query support across unified, RDF, and Property Graph systems.

As shown in the comparison table, the proposed system is the only one to support both RDF and Property Graph storage simultaneously. Furthermore, it provides broader query support, including Basic Graph Patterns (BGP), Text Search, Graph Analysis, and Regular Path Queries (RPQ). Notably, gStore lacks Property Graph support and RPQ, while Neo4j lacks RDF support and RPQ. Experimental results indicate that this unified approach outperforms both baselines in terms of efficiency while maintaining superior feature depth.

Conclusion and Future Work

By utilizing relational algebra as a lingua franca, this system enables interchangeable operations on KGs. A user can perform complex path analysis via Cypher on data originally imported as RDF triples.

Future work will focus on transitioning the system to a distributed architecture to handle large-scale KG data. Additionally, the authors plan to expand the library of supported query features to match the full complexity of the latest SPARQL and Cypher specifications and explore the integration of automated reasoning (RDFS/OWL) into the unified model.

Summary updated

BibTeX
@inproceedings{liu2021towards,
  author = {Baozhu Liu and Xin Wang and Pengkai Liu and Sizhuo Li},
  title = {{Towards a Unified Knowledge Graph Data Management System}},
  booktitle = {SEA-Data@VLDB},
  year = {2021},
  url = {https://openreview.net/forum?id=BHw7XP34cjH},
}
Figure