One Model, Any Conjunctive Query: Graph Neural Networks for Answering Queries over Incomplete Knowledge Graphs (2409.13959v2)

Abstract: Motivated by the incompleteness of modern knowledge graphs, a new setup for query answering has emerged, where the goal is to predict answers that do not necessarily appear in the knowledge graph, but are present in its completion. In this paper, we formally introduce and study two query answering problems, namely, query answer classification and query answer retrieval. To solve these problems, we propose AnyCQ, a model that can classify answers to any conjunctive query on any knowledge graph. At the core of our framework lies a graph neural network trained using a reinforcement learning objective to answer Boolean queries. Trained only on simple, small instances, AnyCQ generalizes to large queries of arbitrary structure, reliably classifying and retrieving answers to queries that existing approaches fail to handle. This is empirically validated through our newly proposed, challenging benchmarks. Finally, we empirically show that AnyCQ can effectively transfer to completely novel knowledge graphs when equipped with an appropriate link prediction model, highlighting its potential for querying incomplete data.

• The paper proposes ANYCQ, a neuro-symbolic GNN framework for classifying and retrieving answers to conjunctive queries over incomplete knowledge graphs.
• It leverages message-passing and reinforcement learning to generalize from simple to complex queries, ensuring robust reasoning even with missing data.
• Experimental results demonstrate superior performance over traditional SQL and existing models, highlighting its scalability and practical impact on real-world data.

One Model, Any Conjunctive Query: Graph Neural Networks for Answering Complex Queries over Knowledge Graphs

Introduction

Knowledge graphs (KGs) have become foundational elements in contemporary data management systems, adeptly handling and representing complex relational data structures. However, real-world KGs often suffer from incompleteness, creating hurdles in retrieving accurate and comprehensive query responses. Traditional methods that rely on a closed-world assumption inadequately manage these gaps, prompting the need for approaches that facilitate reasoning over incomplete datasets through open-world assumptions. The paper presents $\anycq$, a graph neural network (GNN) framework designed to address these challenges by effectively answering conjunctive queries over incomplete knowledge graphs (KGs).

Methodology

The central contribution of the paper is the development of the $\anycq$ GNN model, which operates within a neuro-symbolic framework. This model distinguishes itself by classifying and retrieving answers to Boolean conjunctive queries across any KG, emphasizing both scalability and generalizability. The $\anycq$ framework is specifically crafted to support two query answering tasks:

1. Query Answer Classification (QAC): The model classifies potential answers to a given query as either $\mathsf{true}$ or $\mathsf{false}$.
2. Query Answer Retrieval (QAR): The system either identifies a valid solution or confidently asserts the absence of one.

Query Representation:

• Queries are converted into computational graph structures using a method derived from the ANYCSP framework, defining entities, value vertices, and literals as graph nodes, and distinguishing between entity-value edges and value-literal edges by their respective labels.
• Potential Edge (PE) labels facilitate evaluating feasibility, while Light Edge (LE) labels guide the search for valid assignments during computational processing.

Model Execution:

• An $\anycq$ model searches for the optimal assignment to existential variables over a possible conjunctive Boolean query graph. It employs the hidden states of nodes and informs their evolvement through message-passing techniques within the GNN.

Training and Generalization:

• $\anycq$ is trained using reinforcement learning, allowing it to extrapolate from smaller instance trials to larger, more complex queries. This framework proves valuable in its demonstrated ability to handle queries extending well beyond those it was trained on.

Experimental Evaluation

The authors validate $\anycq$'s efficacy through extensive empirical evaluations across both proposed QAC and QAR benchmarks:

• QAC Performance: When compared against existing query evaluation techniques such as QTO and FIT, $\anycq$ offers on-par performance on simple queries and superior results on complex queries, indicating its robustness in handling various question structures.
• QAR Performance: The model achieves significant success in retrieving both known and unobserved truths, showcasing its ability to extrapolate missing data, distinctly outperforming classical SQL-based query solvers that rely on closed-world assumptions.

  Figure 1: Examples of query graphs of formulas from our FB15k-237-QAR benchmark. Blue nodes represent constants, grey nodes are distinct existentially quantified variables, and orange nodes are free variables.

Implications and Future Directions

The $\anycq$ framework highlights the potential for GNNs to serve as universal engines for complex query answering over knowledge graphs, handling both conjunctive and disjunctive normal form queries. These findings possess implications for the future landscape of AI-driven data management, particularly in domains requiring rapid reasoning over large, incomplete datasets. Future research may look to:

• Enhance $\anycq$'s adaptability to different types of KGs, including hyper-relational and inductively learned graphs.
• Investigate the integration of dynamic knowledge expansion mechanisms to further capture the real-time evolution of complex data networks.
• Extend methodologies to tackle even higher arity queries and further optimize the balance between computational efficiency and accuracy.

In conclusion, $\anycq$ represents a significant stride toward more flexible, accurate query answering systems capable of overcoming traditional data incompleteness obstacles in modern KGs. Its robust generalization capabilities and competitive performance metrics signal a new frontier for scalable query processing technologies.