Efficient k-NN Search on Streaming Data Series

Data streams are common in many recent applications, e.g. stock quotes, e-commerce data, system logs, network traffic management, etc. Compared with traditional databases, streaming databases pose new challenges for query processing due to the streaming nature of data which constantly changes over time. Index structures have been effectively employed in traditional databases to improve the query performance. Index building time is not of particular interest in static databases because it can easily be amortized with the performance gains in the query time. However, because of the dynamic nature, index building time in streaming databases should be negligibly small in order to be successfully used in continuous query processing. In this paper, we propose efficient index structures and algorithms for various models of k nearest neighbor (k-NN) queries on multiple data streams. We find scalar quantization as a natural choice for data streams and propose index structures, called VA-Stream and VA-Stream, which are built by dynamically quantizing the incoming dimensions. VA-Stream (and VA-Stream) can be used both as a dynamic summary of the database and as an index structure to facilitate efficient similarity query processing. The proposed techniques are update-efficient and dynamic adaptations of VA-file (vectorapproximation file) and VA-file, and are shown to achieve the same structures as their static versions. They can be generalized to handle aged queries, which are often used in trend-related analysis. A performance evaluation on VA-Stream and VA-Stream shows that the index building time is negligibly small while query time is significantly improved.