A novel join technique for similar-trend searches supporting normalization on time-series databases

초록

A time-series is defined to be a real-number sequence that is monitored in accordance with a particular time interval. To index a large volume of time-series data without excessive dimensionality expansions, the DFT (Discrete Fourier Transform) technique is widely accepted. It is a challenging task to support fast similarity searches on normalized time-series without false dismissals. Here, the normalization pre-processing on time-series is vital for similar-trend searches that are tackled in our work. To address this problem, we locate multiple sub-queries within a given user query, and map them into points in the normalized DFT index space. Then, a joinlike operation is executed using those points and newly computed Euclidian (similarity) distances. We propose a new cost function utilized for deciding sub-queries that may have the smallest intersection in the index space. With this approach, we can enhance the query performance significantly. Through performance evaluation, it is verified that our approach can reduce the query processing time by about 62%, compared to existing one.

키워드