Scalable Hypergraph Learning and Processing

Jin Huang, Rui Zhang, Jeffrey Xu Yu

Abstract: A hypergraph allows a hyperedge to connect more than two vertices, using which to capture the high-order relationships, many hypergraph learning algorithms are shown highly effective in various applications. When learning large hypergraphs, converting them to graphs to employ the distributed graph frameworks is a common approach, yet it results in major efficiency drawbacks including an inflated problem size, the excessive replicas, and the unbalanced workloads. To avoid such drawbacks, we take a different approach and propose HyperX, which is a thin layer built upon Spark. To preserve the problem size, HyperX directly operates on a distributed hypergraph. To reduce the replicas, HyperX replicates the vertices but not the hyperedges. To balance the workloads, we investigate the hypergraph partitioning problem aiming at minimizing the space and the communication cost subject to two separate constraints on the hyperedge and the vertex workloads. With experiments on both real and synthetic datasets, we verify that HyperX significantly improves the efficiency of the learning algorithms when compared with the graph conversion approach.