A*Net: A Scalable Path-based Reasoning Approach for Knowledge Graphs

Abstract: Reasoning on large-scale knowledge graphs has been long dominated by embedding methods. While path-based methods possess the inductive capacity that embeddings lack, their scalability is limited by the exponential number of paths. Here we present A*Net, a scalable path-based method for knowledge graph reasoning. Inspired by the A* algorithm for shortest path problems, our A*Net learns a priority function to select important nodes and edges at each iteration, to reduce time and memory footprint for both training and inference. The ratio of selected nodes and edges can be specified to trade off between performance and efficiency. Experiments on both transductive and inductive knowledge graph reasoning benchmarks show that A*Net achieves competitive performance with existing state-of-the-art path-based methods, while merely visiting 10% nodes and 10% edges at each iteration. On a million-scale dataset ogbl-wikikg2, A*Net not only achieves a new state-of-the-art result, but also converges faster than embedding methods. A*Net is the first path-based method for knowledge graph reasoning at such scale.

• The paper introduces A*Net, which employs a novel path selection mechanism inspired by the A* algorithm to efficiently reason over large-scale knowledge graphs.
• It uses a neural priority function to learn and select influential nodes and edges, reducing computational complexity while maintaining high performance.
• Experimental results demonstrate A*Net achieves state-of-the-art results on both transductive and inductive tasks, processing only 10% of graph elements.

A*Net: A Scalable Path-based Reasoning Approach for Knowledge Graphs

The paper introduces A*Net, a scalable path-based reasoning model tailored for large-scale knowledge graphs. The complexity of knowledge graphs, derived from vast networks of entities and relations, demands efficient reasoning methods. Traditional embedding methods, despite their scalability, lack inductive reasoning capabilities since they rely heavily on known embeddings for entities. Conversely, path-based methods offer strong inductive potential by leveraging paths between entities. However, they face a scalability bottleneck due to the exponential growth of possible paths. A*Net addresses these challenges by integrating a path-selection mechanism inspired by the A* algorithm.

A*Net distinguishes itself for its ability to selectively identify paths critical for reasoning in knowledge graphs, thereby reducing computational demands. This selection is driven by a learnt priority function that assesses nodes and edges in the graph, reminiscent of the heuristic in the conventional A* algorithm used for shortest path problems. By focusing on a subset of important paths, A*Net scales efficiently even with large datasets.

Methodology

A*Net's methodology revolves around three core processes: path exploration, priority function learning, and iterative reasoning.

• Path Exploration: Inspired by the premise that not all paths contribute equally to reasoning tasks, A*Net focuses on exploring paths that are pertinent to a given query.
• Priority Function Learning: At the heart of A*Net is its priority function, which evaluates the significance of nodes and edges in the context of the query. This neural priority function is trained end-to-end, optimizing it to recognize paths that contribute meaningfully to the reasoning task without explicit hand-crafted heuristics.
• Iterative Reasoning: Mimicking the A* algorithm's iterative refinement, A*Net refines its path selections and reasoning iteratively, ensuring that the reasoning process benefits from accumulated knowledge of path importance.

Experimental Results

The performance of A*Net is evaluated on standard datasets for both transductive and inductive reasoning tasks, with results indicating its competitive performance against established path-based methods like NBFNet on knowledge graphs such as FB15k-237 and WN18RR. Remarkably, A*Net achieves these results by processing only 10% of nodes and edges, striking an optimum balance between performance and computational efficiency.

On the million-scale dataset ogbl-wikikg2, A*Net not only establishes new state-of-the-art results but also demonstrates rapid convergence, outperforming traditional embedding methods. Its inherently distributed and sparse reasoning process allows it to handle datasets significantly larger in scale, a feat unattainable by previous path-based methods.

The paper's results indicate that A*Net also excels in inductive reasoning tasks, generalizing effectively to unseen data. This inductive capability underscores the practical utility of path-based methods, often challenged in large-scale scenarios due to their complexity.

Implications and Future Work

A*Net's success reaffirms the potential of path-based methods in knowledge graph reasoning. Its capability to prioritize and scale path search could catalyze advances in real-world applications like recommendation systems, question answering, and automated deduction systems.

Future research directions could explore the refinement of the priority function to further enhance the precision of path selection, potentially integrating multi-modal data for richer knowledge representations. Moreover, while the current implementation of A*Net demonstrates significant efficiency improvements, future work may focus on system-level innovations that enhance parallel processing capabilities, further optimizing performance benefits in large-scale environments.

In conclusion, A*Net represents a meaningful step forward in path-based reasoning on knowledge graphs, underscoring the viability of integrating intelligent search techniques for scalable and interpretable AI solutions.