Multi-variate Time-Series Feature Selection via Distance Matrix Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning techniques are inadequate for feature selection and dimensionality reduction in multi-variate time-series data, particularly in high-dimensional scenarios, as they fail to effectively combine features and are computationally prohibitive, with existing methods like PCA having high complexity and interpretability issues.
Innovation Solution
The proposed method involves preprocessing multi-variate time-series data to account for sampling intervals and missing data, determining a distance matrix to estimate feature correlation, performing clustering, and reducing dimensionality using techniques like Principal Component Analysis and DBSCAN, enabling the generation of a lower-dimensionality time-series that preserves information and improves computational efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If Principal Component Analysis (PCA) is used for dimensionality reduction, then information preservation is improved, but computational complexity increases exponentially
Solution Approach 1:
The patent segments the dimensionality reduction process into two distinct phases: (1) feature selection phase that identifies and selects relevant features without exponential complexity, and (2) dimensionality reduction phase that applies PCA only on the reduced feature set. This segmentation avoids applying PCA directly on the full high-dimensional dataset, thereby reducing computational complexity while preserving information through the combined approach of feature selection followed by PCA on selected features.
2Adaptability or versatility
If the number of features is increased to capture more patterns, then pattern discovery capability is improved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary feature selection before applying dimensionality reduction techniques. By pre-identifying and selecting the most relevant features using statistical measures and correlation analysis, the system prepares a reduced feature set that maintains pattern discovery capability while significantly reducing the computational burden of subsequent PCA and clustering operations. This preliminary action ensures that only meaningful features are processed in the computationally intensive phases.
3Device complexity
If conventional feature selection methods are used, then computational complexity is reduced, but ability to combine features is lost
Solution Approach 1:
The patent merges multiple feature selection approaches and combines them with dimensionality reduction techniques. Specifically, it combines statistical feature selection (based on correlation with target variable) with PCA-based dimensionality reduction, and further integrates clustering results to identify groups of related features. This merging allows the system to both select individual relevant features and combine correlated features into principal components, achieving both computational efficiency and feature combination capability.
Data Source
AI summary
Systems and methods of automated feature selection and pattern discovery of multi-variate time-series include obtaining a multi-variate times-series from a network; preprocessing the multi-variate times-series to account for sampling intervals and missing data in the multi-variate times-series; determining a distance matrix for the multi-variate times-series which estimates correlation among features in the multi-variate times-series; performing clustering on the distance matrix; reducing dimensionality of the multi-variate times-series based on the clustering to provide a lower-dimensionality time-series; and providing the lower-dimensionality time-series to one or more applications configured to analyze the multi-variate times-series from the network, wherein the lower-dimensionality time-series provides similar information as the multi-variate time-series with fewer dimensions thereby improving computational complexity of the one or more applications.