Multivariate Time Series Segmentation via Sparse Graph Recovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems are inadequate, inefficient, and inflexible in segmenting multivariate time series data, which is complex and noisy, often requiring quadratic time complexity and missing finer-grained patterns.
Innovation Solution
A time series segmentation system that employs graph recovery models to convert multivariate time series into graph objects and uses similarity models to determine segmentation timestamps, achieving linear time complexity and scalability through batch processing with deep-learning sparse graph recovery models like uGLAD.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional systems are used for multivariate time series segmentation, then they can process the data, but they require quadratic time complexity and are inefficient
Solution Approach 1:
The patent segments the multivariate time series into multiple univariate time series, one for each variable. This allows the system to process each variable independently using efficient univariate segmentation algorithms, avoiding the quadratic time complexity that would result from treating all variables together as a single multivariate sequence.
Solution Approach 2:
The patent transforms the problem from the time dimension alone to include a variable dimension. By organizing data as a matrix where rows represent time points and columns represent variables, the system can process along the time dimension for each variable separately, converting an O(n²) problem into multiple O(m) problems where m is the number of variables.
2Adaptability or versatility
If conventional systems focus on univariate cases, then they are simpler to implement, but they are inadequate for multivariate time series data
Solution Approach 1:
The patent creates a universal segmentation system that can handle both univariate and multivariate time series data. The same core segmentation algorithm is applied to each variable independently, allowing the system to process any combination of variables without requiring different algorithms or increasing complexity. This multi-functional approach maintains simplicity while expanding capability.
3Measurement precision
If conventional systems segment time series, then they provide basic segmentation, but they miss finer-grained patterns and are inflexible
Solution Approach 1:
The patent applies different segmentation parameters and window sizes to different variables based on their individual characteristics. Each variable can have its own optimal segmentation settings, allowing the system to capture finer-grained patterns specific to each variable while maintaining overall system simplicity through the use of standard segmentation algorithms applied locally to each variable.
Data Source
AI summary
This disclosure relates to a time series segmentation system that automatically segments multivariate time series data. For example, the time series segmentation system is capable of converting complex and noisy multivariate time series data into segmented multivariate time series by identifying distinct segments within the data. The time series segmentation system operates with linear time complexity in terms of sequence length, which is significantly more efficient than the typical quadratic time complexity required by conventional systems. To illustrate, the time series segmentation system first divides a multivariate time series into portions using time-based windows. The time series segmentation system then converts the windowed subsequences into graph objects using a sparse graph recovery model and utilizes a similarity model to determine segmentation timestamps from the graph objects. The time series segmentation system then uses the segmentation timestamps to convert the multivariate time series data into a segmented multivariate time series.