Trading Quality for Efficiency of Graph Partitioning: An Inductive Method across Graphs

Many applications of network systems can be formulated as several NP-hard combinatorial optimization problems regarding graph partitioning (GP), e.g., modularity maximization and NCut minimization. Due to the NP-hardness, to balance the quality and efficiency of GP remains a challenge. Existing methods use machine learning techniques to obtain high-quality solutions but usually have high complexity. Some fast GP methods adopt heuristic strategies to ensure low runtime but suffer from quality degradation. In contrast to conventional transductive GP methods applied to a static graph, we propose an inductive graph partitioning (IGP) framework across multiple evolving graph snapshots to alleviate the NP-hard challenge. IGP first conducts the offline training of a novel dual graph neural network on historical snapshots to capture the structural properties of a system. The trained model is then generalized to newly generated snapshots for fast high-quality online GP without additional optimization, where a better trade-off between quality and efficiency is achieved. IGP is also a generic framework that can capture the permutation invariant partitioning ground-truth of historical snapshots in the offline training and tackle the online GP on graphs with non-fixed number of nodes and clusters. Experiments on a set of benchmarks demonstrate that IGP achieves competitive quality and efficiency to various state-of-the-art baselines.