Faster way to rank important points in networks over time

New parallel algorithms compute the higher-order temporal H-index much faster than previous methods.

Researchers developed parallel algorithms for higher-order temporal H-index computation. This method ranks vertices by importance in networks.

The H-index ranks vertices by how many neighbors are themselves well connected. The higher-order temporal variant applies this idea recursively over time-respecting neighborhoods in temporal networks.

Existing streaming algorithms are difficult to scale. They intertwine temporal processing with H-index order, which incurs maximum-degree-dependent costs and limits parallelism.

The new approach resolves temporal dependencies once. Each edge's time-dependent lookup maps to a fixed successor position. This turns every H-index order into a parallel gather followed by a segmented suffix-H computation.

The researchers created a linear-work suffix-H algorithm and a multicore CPU implementation. They also built degree-adaptive GPU kernels and a multi-GPU scheme.

On a 636M-edge Reddit network, eight GPUs computed 128 orders in 3.6 s of round time. The streaming algorithm did not finish within an hour even at order 8.

Key facts

  • The new reformulation is substantially faster than the streaming algorithm.
  • GPUs provide speedups exceeding an order of magnitude over multicore implementations on synthetic inputs with up to one billion temporal edges.
  • The multi-GPU scheme exchanges only distinct remotely queried states.

All sections

Edition