Learning-Based Dynamic Graph Stream Sketch

Abstract

A graph stream is a kind of dynamic graph representation that consists of a consecutive sequence of edges where each edge is represented by two endpoints and a weight. Graph stream is widely applied in many application scenarios to describe the relationships in social networks, communication networks, academic collaboration networks, etc. Graph sketch mechanisms were proposed to summarize large-scale graphs by compact data structures with hash functions to support fast queries in a graph stream. However, the existing graph sketches use fixed-size memory and inevitably suffer from dramatic performance drops after a massive number of edge updates. In this paper, we propose a novel Dynamic Graph Sketch (DGS) mechanism, which is able to adaptively extend graph sketch size to mitigate the performance degradation caused by memory overload. The proposed DGS mechanism incorporates deep neural network structures with graph sketch to actively detect the query errors, and dynamically expand the memory size and hash space of a graph sketch to keep the error below a pre-defined threshold. We conducted extensive experiments on three real-world graph stream datasets, which show that DGS outperforms the state-of-the-arts with regard to the accuracy of different kinds of graph queries.