Primal-Dual Neural Algorithmic Reasoning

2025-05-29

AI Summary

This work extends Neural Algorithmic Reasoning (NAR) to NP-hard combinatorial optimization problems, addressing the limitation of existing NAR approaches—restricted to polynomial-time solvable tasks. We propose the first neural framework deeply integrating the primal-dual approximation algorithm paradigm: a bipartite graph representation explicitly models interactions between primal and dual variables, and optimal solution distillation from small-scale instances enhances generalization. The method synergistically combines graph neural networks, primal-dual theory, supervised algorithmic imitation, and co-training with commercial solvers. Experiments demonstrate that our approach significantly outperforms classical approximation algorithms across multiple NP-hard benchmarks. It exhibits strong generalization on large-scale graphs and out-of-distribution instances, while maintaining practical utility and interpretability—as validated on real-world datasets.

Technology Category

• Search and Optimization: Combinatorial Optimization
• Constraint Satisfaction and Optimization: Other Foundations of Constraint Satisfaction
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Neural Algorithmic Reasoning (NAR) trains neural networks to simulate classical algorithms, enabling structured and interpretable reasoning over complex data. While prior research has predominantly focused on learning exact algorithms for polynomial-time-solvable problems, extending NAR to harder problems remains an open challenge. In this work, we introduce a general NAR framework grounded in the primal-dual paradigm, a classical method for designing efficient approximation algorithms. By leveraging a bipartite representation between primal and dual variables, we establish an alignment between primal-dual algorithms and Graph Neural Networks. Furthermore, we incorporate optimal solutions from small instances to greatly enhance the model's reasoning capabilities. Our empirical results demonstrate that our model not only simulates but also outperforms approximation algorithms for multiple tasks, exhibiting robust generalization to larger and out-of-distribution graphs. Moreover, we highlight the framework's practical utility by integrating it with commercial solvers and applying it to real-world datasets.

Problem

• Extending Neural Algorithmic Reasoning to harder problems
• Aligning primal-dual algorithms with Graph Neural Networks
• Enhancing reasoning capabilities using small instance solutions

Innovation

• Primal-dual paradigm enhances neural algorithmic reasoning
• Bipartite representation aligns with Graph Neural Networks
• Incorporates small instance solutions to boost reasoning