Random Time Warping for Time-Series Feature Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning methods face challenges in processing time-series data due to its variable length and dynamic time scaling, as existing techniques like DTW-based methods suffer from high computational complexity and memory requirements, and previous global alignment kernels are impaired by diagonal dominance and scalability issues.
Innovation Solution
The approach generates a set of reference time-series of random lengths, transforming raw time-series data into a low-dimensional feature space using random warping series, which reduces computational complexity from O(N2L2) to O(NRLD) and memory consumption from O(NL+N2) to O(NR), allowing for more efficient processing and training with machine learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If DTW-based methods are used for time-series processing, then alignment accuracy is improved, but computational complexity increases to O(N2L2)
Solution Approach 1:
The patent segments the time-series alignment problem by introducing random warping series that divide the alignment task into multiple random projections. Instead of computing all-to-all DTW alignments between N time-series of length L, the method projects each time-series onto R random warping series, reducing complexity from O(N2L2) to O(NRLD) where R is the number of random series and D is their dimensionality.
Solution Approach 2:
The patent introduces random warping series as intermediary objects that mediate the alignment process. These random series serve as intermediate representations that capture temporal relationships without requiring direct pairwise comparison of all time-series points, thus reducing computational burden while preserving alignment accuracy.
2Measurement precision
If global alignment kernels are used for time-series analysis, then comprehensive comparison is improved, but scalability deteriorates due to diagonal dominance
Solution Approach 1:
The patent replaces static global alignment kernels with dynamic random warping series that adapt to different time-series characteristics. The random warping series are generated dynamically and can capture varying temporal patterns without suffering from the diagonal dominance problem that plagues global alignment kernels, enabling better scalability.
Solution Approach 2:
The patent changes the fundamental parameters of the alignment approach by transitioning from fixed kernel parameters to random warping parameters. This parameter change allows the system to maintain comprehensive comparison capabilities while avoiding the scalability limitations of traditional global alignment kernels through randomized parameter selection.
3Reliability
If traditional time-series methods are used, then data fidelity is preserved, but memory consumption increases to O(NL+N2)
Solution Approach 1:
The patent extracts only the essential temporal features by projecting time-series onto random warping series, discarding redundant information. This extraction process reduces memory consumption from O(NL+N2) to O(NR) by storing only the projected features rather than the full time-series data and kernel matrices, while preserving the critical temporal relationships needed for accurate analysis.
Data Source
AI summary
Embodiments of the present invention provide a computer-implemented method for performing unsupervised time-series feature learning. The method generates a set of reference time-series of random lengths, in which each length is uniformly sampled from a predetermined minimum length to a predetermined maximum length, and in which values of each reference time-series in the set are drawn from a distribution. The method generates a feature matrix for raw time-series data based on a set of computed distances between the generated set of reference time-series and the raw time-series data. The method provides the feature matrix as an input to one or more machine learning models.