Our mission is to build the most scalable library for graph algorithms and analysis and apply it to a multitude of Google products.
About the team
We formalize data mining and machine learning challenges as graph problems and perform fundamental research in those fields leading to publications in top venues. Our algorithms and systems are used in a wide array of Google products such as Search, YouTube, AdWords, Play, Maps, and Social.
Team focus summaries
Large-Scale Balanced Partitioning
Balanced Partitioning splits a large graph into roughly equal parts while minimizing cut size. The problem of “fairly” dividing a graph occurs in a number of contexts, such as assigning work in a distributed processing environment. Our techniques provided a 40% drop in multi-shard queries in Google Maps driving directions, saving a significant amount of CPU usage.
Our team specializes in clustering graphs at Google scale, efficiently implementing many different algorithms including hierarchical clustering, overlapping clustering, local clustering, and spectral clustering.
Large-Scale Connected Components
Connected Components is a fundamental subroutine in many graph algorithms. We have state-of-the-art implementations in a variety of paradigms including MapReduce, a distributed hash table, Pregel, and ASYMP. Our methods are 10-30x faster than the best previously studied algorithms, and easily scale to graphs with trillions of edges.
Large-Scale Similarity Ranking
Public-private Graph Computation
Our research on novel models of graph computation addresses important issues of privacy in graph mining. Specifically, we present techniques to efficiently solve graph problems, including computing clustering, centrality scores and shortest path distances for each node, based on its personal view of the private data in the graph while preserving the privacy of each user.
Streaming and Dynamic Graph Algorithms
We perform innovative research analyzing massive dynamic graphs. We have developed efficient algorithms for computing densest subgraph and triangle counting which operate even when subject to high velocity streaming updates.
ASYMP: Async Message Passing Graph Mining
ASYMP is a graph mining framework based on asynchronous message passing. We have highly scalable code for Connected Components and shortest-path to a subset of nodes in this framework.
Large-Scale Centrality Ranking
Google’s most famous algorithm, PageRank, is a method for computing importance scores for vertices of a directed graph. In addition to PageRank, we have scalable implementations of several other centrality scores, such as harmonic centrality.
Adaptive Massively Parallel Computation (AMPC) augments the theoretical capabilities of MapReduce, providing a pathway to solve many graph problems in fewer computation rounds; the suite of algorithms, which includes algorithms for maximal independent set, maximum matching, connected components and minimum spanning tree, work up to 7x faster than current state-of-the-art approaches.
Solving large-scale optimization problems often starts with graph partitioning, which means partitioning the vertices of the graph into clusters to be processed on different machines. The need to make sure that clusters are of near equal size gives rise to the balanced graph partitioning problem. This NP-hard problem is notoriously difficult in practice because the best approximation algorithms for small instances rely on semidefinite programming which is impractical for larger instances.
We present the results of two recent papers on graph embedding: “Is a Single Embedding Enough? Learning Node Representations that Capture Multiple Social Contexts'' presented at WWW’19 and “Watch Your Step: Learning Node Embeddings via Graph Attention” at NeurIPS’18. The first paper introduces a novel technique to learn multiple embeddings per node, enabling a better characterization of networks with overlapping communities. The second addresses the fundamental problem of hyperparameter tuning in graph embeddings, allowing one to easily deploy graph embedding methods with less effort.
The Mining and Learning with Graphs at Scale workshop focused on methods for operating on massive information networks: graph-based learning and graph algorithms for a wide range of areas such as detecting fraud and abuse, query clustering and duplication detection, image and multi-modal data analysis, privacy-respecting data mining and recommendation, and experimental design under interference.