Graph-based approaches for Analysing and Accelerating Streaming Data
主 题: Graph-based approaches for Analysing and Accelerating Streaming Data
报告人: Prof. Weiru Liu (Queen s University Belfast, UK))
时 间: 2016-09-13 10:00-11:00
地 点: 理科1号楼1556
Graph-based approaches are an powerful way to analyse data, especially when items in data with natural connections, such as social networks, sensor networks, biological networks etc. In this talk, we present two graph-based methods. The first method is for analyzing streaming data utilizing the count-min idea in order to deal with unbounded data streams. In this method, each graph data item is extended to include additional (side) information in order to make the clustering more accurate. The second method is about accelerating large scale centroid-based clustering (both graph and non-graph) algorithms with locality sensitive hashing. Two-layer multiple hash functions are used to generate candidate clusters so as to speed up the algorithms up to factor of 6. Both algorithms have been extensively evaluated with real, artificial and benchmark datasets. Finally, if time permits, I will briefly introduce our novel approach for handling both numerical and structural anomalies in graph-based data analytics for anomaly detection.