Time-Series Pattern Extraction for Constant and Random Behavior
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques fail to effectively distinguish and extract patterns related to behavior with temporal constancy from those without, leading to misclassification of random behavior as noise, especially when such behavior occurs frequently or with large observation values.
Innovation Solution
An information processing device constructs secondary data by stacking primary time-series data and uses tensor factorization with constraints like Group Lasso regularization to extract constant patterns, then generates differences to isolate random patterns without temporal constancy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If matrix decomposition techniques (PCA or SVD) are used to detect main patterns in data, then the analysis can be performed efficiently, but the behavior having temporal constancy and random behavior cannot be distinguished from each other, and random behavior may be regarded as noise
Solution Approach 1:
The patent segments the pattern detection process into two distinct stages: first extracting behavior patterns with temporal constancy using tensor decomposition, then extracting random behavior patterns from the residual data. This segmentation allows each extraction process to focus on specific pattern types, preventing random behavior from being misclassified as noise while maintaining efficient analysis through structured processing steps.
Solution Approach 2:
The patent introduces a temporal constancy dimension by stacking data along the time axis to create tensor structures. This additional dimensional organization enables the decomposition process to identify and separate patterns based on their temporal characteristics, distinguishing between behaviors that repeat over time and those that occur randomly, thereby improving pattern distinction accuracy.
2Measurement precision
If tensor decomposition is applied to stacked data to extract periodic patterns, then behavior having temporal constancy can be detected, but random behavior emerging frequently or with large observation values may be mixed with periodic behavior patterns
Solution Approach 1:
The patent extracts periodic behavior patterns first through tensor decomposition of stacked data, then separately extracts random behavior patterns from the residual data that remains after removing periodic components. This sequential extraction ensures that random behavior information is not lost or mixed with periodic patterns, as each type is isolated and processed independently in its own extraction stage.
Solution Approach 2:
The patent employs a dynamic two-stage extraction process where the focus shifts from extracting periodic patterns in the first stage to extracting random patterns in the second stage. This dynamic approach adapts to different pattern types at different processing stages, allowing frequent or large-magnitude random behaviors to be captured separately rather than being conflated with periodic behavior.
3Measurement precision
If data is stacked in time direction to generate tensor with redundancy, then periodic patterns can be easily extracted, but the calculation amount becomes large and results may not converge
Solution Approach 1:
The patent segments the overall pattern extraction task into two separate decomposition processes: first decomposing the stacked tensor to extract periodic patterns, then decomposing the residual data to extract random patterns. This segmentation reduces the complexity of each individual decomposition operation compared to attempting to extract all patterns simultaneously, while maintaining the benefits of temporal redundancy for periodic pattern detection.
Data Source
AI summary
An information processing device according to the present invention includes: a memory; and at least one processor coupled to the memory. The processor performing operations. The operations includes: constructing second data that is acquired, based on first data containing a plurality of observation values in a plurality of times, by stacking the first data with respect to the times, and extracting a constant pattern that is a combination of the observation values having temporal constancy in the first data, based on the second data; generating a difference between the first data and the constant pattern in the time; and extracting a random pattern that is a combination of the observation values without temporal constancy, based on the difference.