Trend-Based Similarity Search in Time-Series Data

Abstract

In this paper, we present a novel approach towards time-series similarity search. Our technique relies on trends in a curve’s movement over time. A trend is characterized by a series’, values channeling in a certain direction (up, down, sideways) over a given time period before changing direction. We extract trend-turning points and utilize them for computing the similarity of two series based on the slopes between their turning points. For the turning point extraction, well-known techniques from financial market analysis are applied. The method supports queries of variable lengths and is resistant to different scaling of query and candidate sequence. It supports both subsequence searching and full sequence matching. One particular focus of this work is to enable simple modeling of query patterns as well as efficient similarity score updates in case of appending new data points.