Similarity search over time series and trajectory data

• Authors: Lei Chen
• Venue: Similarity search over time series and trajectory data
• Year: 2005

Time series data have been used in many applications, such as financial data analysis and weather forecasting. Similarly, trajectories of moving objects are often used to perform movement pattern analysis in surveillance video and sensor monitoring systems. All these applications are closely related to similarity-based time series or trajectory data retrieval. In this dissertation, various similarity models are proposed to capture the similarities among time series and trajectory data under various circumstances and requirements, such as the appearance of noise and local time shifting. A novel representation, called multi-scale time series histograms , is proposed to answer pattern existence queries and shape match queries. Earlier proposals generally address one or the other; multi-scale time series histograms can answer both types, which offers users more flexibility. A metric distance function, called Edit distance with Real Penalty (ERP), is proposed that can support local time shifting in time series and trajectory data. A second distance function, Edit Distance on Real sequence (EDR) is proposed to measure the similarity between time series or trajectories with local time shifting and noise. Since the proposed similarity models are computationally expensive, several indexing and pruning methods are proposed to improve the retrieval efficiency. For multi-scale time series histograms, A multi-step filtering process is introduced to improve the retrieval efficiency without introducing false dismissals. For ERP, a framework is developed to index time series or trajectory data under a metric distance function, which exploits the pruning power of lower bounding and triangle inequality. For EDR, three pruning techniques—mean value Q-grams, near triangle inequality, and trajectory histograms—are developed to improve the retrieval efficiency.