Homogeneous Network Embedding for Massive Graphs via Personalized PageRank

Given an input graph G and a node v in G, homogeneous network embedding (HNE) aims to map the graph structure in the vicinity of v to a fixed-dimensional feature vector. Such feature vectors can then be fed to a machine learning pipeline as inputs, e.g., for link prediction. The vast majority of existing approaches to HNE are immensely expensive, and, thus, are limited to small graphs. The few that do scale to large graphs, on the other hand, are based on quick-and-rough heuristics that fail to fully capture the local graph structure, leading to limited effectiveness of the extracted feature vectors. Meanwhile, in recent years there have been rapid advancements in scalable algorithms for computing approximate personalized PageRank (PPR), which captures rich graph topological information. A natural idea is: can we construct an effective embedding of a node v based on its approximate PPR values? This turns out to be challenging, due to the fact that PPR was designed for a very different purpose, i.e., ranking nodes in the graph based on their relative importance from an individual node v's perspective. In contrast, in HNE the goal is to build unified node embeddings for all nodes from a global view. Consequently, using PPR values directly as node embeddings leads to sub-optimal result effectiveness. Motivated by this, we propose Node-Reweighted PageRank (NRP), a novel solution that transforms personalized PageRank values into effective node embeddings, by iteratively solving an optimization problem. NRP computation is efficient and scalable: in particular, NRP embeddings for a billion-edge graph can be computed in a few hours on a single CPU core. At the same time, the quality of NRP embeddings is highly competitive: as our experiments demonstrate, in terms of prediction accuracy, NRP consistently outperforms existing HNE methods on 3 graph analysis tasks, on both large and small graphs.