Synthetic Time Series Change Points for Realistic ML Benchmarking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often lack sufficient and relevant data for training and benchmarking, particularly for detecting specific irregularities such as anomalies and change points, leading to inadequate performance.
Innovation Solution
Generate synthetic time series datasets with labeled anomalies and/or change points using components like seasonality, trend, and noise, ensuring each component contributes appropriately to maintain dataset integrity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If synthetic datasets are generated to increase data availability, then the quantity of training data is improved, but the realism and irregularity of the data may deteriorate
Solution Approach 1:
The synthetic time series data is segmented into multiple independent components (trend component, seasonal component, irregular component) that are generated separately and then combined. This allows each component to be controlled independently to maintain statistical properties while increasing overall data quantity.
Solution Approach 2:
The patent applies parameter changes by transforming the irregular component through scaling factors and combining it with trend and seasonal components using weighted parameters. This maintains the irregularity characteristics while adjusting the overall data distribution to resemble real time series data.
2Measurement precision
If synthetic datasets include labeled irregularities for training anomaly detection models, then the measurement precision of model performance is improved, but the complexity of data generation increases
Solution Approach 1:
The patent applies preliminary action by pre-defining the structure and characteristics of irregularities (anomalies and change points) before generating the synthetic data. The irregular components are prepared in advance with known statistical properties, and labels are automatically assigned based on the generated patterns, enabling precise model benchmarking without complex post-processing.
3Reliability
If multiple data generation functions are combined to create realistic time series data, then the realism of the synthetic data is improved, but the difficulty of controlling component balance increases
Solution Approach 1:
The patent uses parameter changes by introducing scaling factors (e.g., scaling the irregular component by a factor between 0.1 and 0.5) and combining components using weighted sums. These parameters allow controlled adjustment of each component's contribution to the final synthetic data, maintaining realism while enabling measurement and control of component balance.
Data Source
AI summary
Methods and systems are described herein for facilitating generation of synthetic datasets having a change point. The system may receive a command to generate a synthetic time series dataset. The system may generate data points for components of the synthetic dataset, the components including a seasonality function, a trend function, and a noise function. The system may modify the trend function to a different trend function by modifying a level or a slope of the trend function. The system may generate a change point by replacing a subset of consecutive data points generated using the trend function with consecutive data points generated using the different trend function. The system may then generate the synthetic time series dataset having a change point by combining the seasonality data points, the trend data points, and the noise data points into corresponding time slots of the synthetic time series dataset.