Scalable Similarity Search Algorithm for High-dimensional Time Series Data

Abstract

Similarity Search is important for data processing and identifying similar time series is a core subroutine for many data analysis problems. However, searching for similar items faces several challenges. Specifically for time series data, first, searching over complex objects, such as massive multivariate time series database query is computationally expensive in accordance with the alignment-based similarity measurement Dynamic Time Warping (DTW). Second, existing efficient solutions largely focus on univariate time series and few intends to scale as the dimensionality increases. Given these situations, a Locality Sensitive Hashing (LSH)-based fast similarity search method for multivariate time series was previously developed. Moreover, Machine Learning is developing rapidly nowadays. Numerous Deep Learning researches showed promising achievements on various domains, such as tackling classification and clustering problems. And the intermediate results from the Deep Learning process are meaningful with implicit features expressions, which ought to be useful for more data analysis researches. This project provides a solution of combining Deep Learning features extraction mechanism through univariate time series dimension expansions and proposed LSH-based fast similarity search method for multivariate time series, intending to achieve decent results with convincing accuracy and efficiency improvements.