Synthetic Time-Series Dataset Resampling for Rare-Event Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Class imbalance in training datasets, particularly in time-series data, adversely affects the accuracy of machine-learning models by skewing the distribution of labels towards the majority class, leading to poor performance in identifying rare events.
Innovation Solution
Generate synthetic time series for the minority class and/or reduce the number of time series for the majority class using up-sampling and down-sampling techniques that respect the temporal correlation of time samples, thereby balancing the dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the training dataset is used as-is with natural class distribution, then the dataset requires no processing, but the machine-learning model achieves poor accuracy for rare events due to severe class imbalance
Solution Approach 1:
The patent creates synthetic copies of minority class time series by interpolating between neighboring time series in the feature space. These synthetic time series are generated by combining features from multiple neighboring time series using weighted averages, where weights are determined by distance metrics. This copying approach increases the number of minority class samples without requiring actual rare event data, thereby improving model accuracy while maintaining data privacy and avoiding speculative data generation.
Solution Approach 2:
The patent transforms the class distribution parameters of the training dataset by adjusting the representation of minority class samples. Through synthetic minority over-sampling technique (SMOTE), the patent changes the effective class balance parameter from highly skewed to more balanced distribution. This parameter change enables the machine-learning model to learn patterns from rare events more effectively, improving measurement precision for rare event detection.
2Measurement precision
If synthetic time series are generated for minority class using neighboring time series, then class imbalance is reduced and model accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the minority class time series generation process into discrete steps: (1) identifying neighboring time series for each minority class sample, (2) calculating distance metrics to determine weights, (3) generating synthetic samples by combining neighboring features, and (4) adding synthetic data to the training set. This segmentation allows the complex task of synthetic data generation to be broken down into manageable computational operations, reducing overall computational complexity while maintaining accuracy improvements.
Solution Approach 2:
The patent applies partial action by generating synthetic time series only for the minority class rather than processing the entire dataset uniformly. This selective approach focuses computational resources only where needed - on the minority class samples that require augmentation. The excessive action is reflected in generating multiple synthetic samples per minority class representative, which over-samples the minority class to achieve balanced distribution, thereby improving model accuracy without unnecessarily processing majority class data.
3Quantity of substance
If up-sampling and down-sampling techniques are applied to balance classes, then the training dataset becomes balanced, but the temporal correlation of time samples may be compromised
Solution Approach 1:
The patent resolves the contradiction by operating in a transformed feature space rather than directly manipulating time series in the temporal dimension. By projecting time series into a feature space where temporal patterns are captured as feature vectors, and then performing synthetic data generation in this transformed space, the patent maintains temporal correlation structure while achieving class balance. The synthetic time series inherit temporal characteristics from their neighboring counterparts through the interpolation process, preserving temporal correlation despite the artificial creation of samples.
Data Source
AI summary
Class imbalance in a training dataset may negatively impact the accuracy of a machine-learning model in classifying rare events that are underrepresented in the training dataset. Training datasets comprising time-series data present a unique challenge. Accordingly, resampling techniques for up-sampling and/or down-sampling a training dataset of time series are disclosed. The up-sampling may respect the temporal correlation of time samples in the time series, while generating synthetic time series that mimic the feature values of time series belonging to the minority class. Down-sampling may be used to fine-tune the ratio of time series belonging to the minority class to the time series belonging to the majority class.