PSEUDo: Interactive Pattern Search in Multivariate Time Series with Locality-Sensitive Hashing and Relevance Feedback

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• High-dimensional time series data is rapidly increasing due to sensor technology.
• Existing deep learning methods for pattern retrieval are computationally expensive and require extensive training data.
• Automatic retrieval methods often lack user-specific similarity notions and ground truth labels.

Purpose of the Study:

• Introduce PSEUDo, a visual pattern retrieval tool for multivariate time series.
• Address the challenge of inefficient retraining in deep learning-based time series analysis.
• Provide a user-friendly and adaptable solution for high-dimensional time series pattern search.

Main Methods:

• Developed PSEUDo based on query-aware locality-sensitive hashing (LSH).
• Employs LSH to create representations of multivariate time series windows.
• Achieves sub-linear training and inference time complexity concerning data dimensions.

Main Results:

• PSEUDo demonstrates efficient pattern detection in high-dimensional time series.
• The tool achieves instantaneous relevance-feedback-driven adaptation to user similarity.
• Performance is validated through quantitative benchmarks and a case study, showing improvements in accuracy, speed, and usability.

Conclusions:

• PSEUDo offers an efficient and user-friendly approach to multivariate time series pattern retrieval.
• The method overcomes the limitations of deep learning retraining costs.
• Relevance feedback enhances retrieval accuracy by enabling feature selection, aligning with user needs.