FeifeiCS/NGDBench

NGDBench是一个用于评估在噪声、不完整和演化观察下的神经图数据管理的基准数据集。它设计用于测试当前系统是否能超越对观测数据的被动查询，并推理出数据所隐含的潜在结构。该数据集涵盖五个领域：社交系统、金融、生物医学、工具使用工作流和企业报告，整合了结构化和非结构化数据源，并以图表示形式呈现。每个结构化数据集都提供干净的潜在图和受现实扰动影响的观察图，支持系统评估噪声鲁棒查询。NGDBench还支持完整的Cypher风格分析工作负载和动态数据管理操作，包括V1.1版本中新增的分析查询、布尔验证查询、管理查询和自然语言查询字段。数据集包含五个子数据集：NGD-BI（社交网络图）、NGD-Fin（金融交易图）、NGD-Prime（生物医学知识图）、NGD-MCP（工具使用工作流图）和NGD-Econ（企业报告图），适用于文本到Cypher生成、图RAG、噪声鲁棒查询回答等研究。NGDBench is a benchmark dataset for evaluating neural graph data management under noisy, incomplete, and evolving observations. It is designed to test whether current systems can go beyond passive querying of observed data and infer the latent structures implicit within the data. This dataset covers five domains: social systems, finance, biomedicine, tool-use workflows, and enterprise reporting. It integrates structured and unstructured data sources and presents them in graph representation. Each structured dataset provides both a clean latent graph and an observed graph affected by realistic perturbations, enabling system evaluation for noise-robust querying. NGDBench also supports full Cypher-style analytical workloads and dynamic data management operations, including newly added analytical queries, boolean validation queries, administrative queries, and natural language query fields in version 1.1. The dataset includes five subsets: NGD-BI (social network graph), NGD-Fin (financial transaction graph), NGD-Prime (biomedical knowledge graph), NGD-MCP (tool-use workflow graph), and NGD-Econ (enterprise reporting graph), which are applicable to research such as text-to-Cypher generation, graph RAG, noise-robust query answering, and other related studies.