Mining Dynamic Graph Streams for Predictive Queries Under Resource Constraints

Knowledge graph streams are a data model underlying many online dynamic data applications today. Answering predictive relationship queries over such a stream is very challenging as the heterogeneous graph streams imply complex topological and temporal correlations of knowledge facts, as well as fast dynamic incoming rates and statistical pattern changes over time. We present our approach with two major components: a Count-Fading sketch and an online incremental embedding algorithm. We answer predictive relationship queries using the embedding results. Extensive experiments over real world datasets show that our approach significantly outperforms two baseline approaches, producing accurate query results efficiently with a small memory footprint.