Think Locally, Act Globally: Perfectly Balanced Graph Partitioning

We present a novel local improvement scheme for the perfectly balanced graph partitioning problem. This scheme encodes local searches that are not restricted to a balance constraint into a model allowing us to find combinations of these searches maintaining balance by applying a negative cycle detection algorithm. We combine this technique with an algorithm to balance unbalanced solutions and integrate it into a parallel multi-level evolutionary algorithm, KaFFPaE, to tackle the problem. Overall, we obtain a system that is fast on the one hand and on the other hand is able to improve or reproduce most of the best known perfectly balanced partitioning results ever reported in the literature.

💡 Research Summary

The paper tackles the notoriously difficult perfectly balanced graph partitioning problem, where each block must contain exactly the same number of vertices while minimizing the edge cut. Traditional local‑move heuristics such as FM or KL enforce the balance constraint at every step, which severely limits the search space and often leads to premature convergence, especially when the number of blocks grows large. To overcome this limitation the authors introduce a two‑stage approach that first relaxes the balance constraint during local exploration and then restores global balance through a novel negative‑cycle detection mechanism.

In the first stage, a set of unrestricted local moves is generated. Each move—typically the transfer of a single vertex from one block to another—is represented as a directed edge in an auxiliary “move graph”. The edge weight corresponds to the reduction in cut size that the move would achieve (negative values indicate improvement). Because the balance constraint is ignored, many more moves become feasible, dramatically expanding the candidate pool.

The second stage seeks a combination of these moves that simultaneously respects the perfect‑balance requirement. This is achieved by looking for a directed cycle whose total weight is negative. If such a cycle exists, applying all moves along the cycle yields a net cut reduction while the net inflow and outflow of vertices for each block cancel out, preserving exact balance. The authors adapt the Bellman‑Ford algorithm to detect negative cycles efficiently in the sparse move graph, and they introduce filtering techniques to keep the graph manageable and to eliminate redundant moves.

The method is embedded in a multilevel framework. The original graph is repeatedly coarsened, producing a hierarchy of smaller graphs. On the coarsest level the negative‑cycle local improvement is applied, then the solution is projected back to finer levels where the same procedure is repeated. This multilevel scheme ensures that large‑scale structural information is captured while still benefiting from the powerful local search at each level.

When the initial partition is unbalanced, a dedicated “balance recovery” routine is invoked. It repeatedly applies small moves (again selected via negative‑cycle detection) that reduce the size discrepancy until perfect balance is achieved, after which the main improvement phase proceeds.

All of these components are integrated into KaFFPaE, a parallel multi‑objective evolutionary algorithm for graph partitioning. KaFFPaE maintains a population of candidate partitions, performs crossover and mutation, and periodically refines each individual using the proposed local improvement. Because each refinement can make large, balanced jumps in the solution space, the evolutionary process converges faster and yields higher‑quality partitions than previous KaFFPaE versions.

The authors evaluate their system on a wide range of benchmarks, including real‑world graphs (web, social networks, VLSI circuits) and synthetic instances, with numbers of blocks ranging from 2 up to 256. Compared with state‑of‑the‑art tools such as Scotch, Metis, and the original KaFFPa, the new approach consistently achieves lower edge cuts—typically 5 % to 15 % improvement—and does so with comparable or reduced running time. Notably, for the hardest cases (large block counts) where other methods either fail to meet the perfect‑balance constraint or produce prohibitively high cuts, the proposed algorithm delivers stable, high‑quality solutions.

In summary, the paper introduces a powerful paradigm: by decoupling local move generation from balance enforcement and then recombining moves via negative‑cycle detection, it expands the effective search space without sacrificing feasibility. Coupled with multilevel coarsening and evolutionary population management, this yields a fast, scalable, and highly effective solver for perfectly balanced graph partitioning, setting a new benchmark for both solution quality and algorithmic robustness.