Lookback Window Parameter Search via Meta-Feature Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated machine learning systems for time series data face challenges in efficiently searching for optimal lookback window parameters, leading to slow pipeline construction and poor model performance due to the significant increase in search space when these parameters change.
Innovation Solution
The method involves creating predefined pipelines with meta-features and segmenting time series data using lookback window parameters to identify those parameters that result in windowed data with meta-features most similar to predefined pipelines, employing techniques like Randomize search and Bayesian optimization to quickly find optimal parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If lookback window parameters are changed to improve model performance, then model accuracy is improved, but the search space significantly increases leading to slower pipeline construction
Solution Approach 1:
The patent pre-calculates and stores meta-features for multiple predefined pipelines before the actual model training process. This preliminary action allows the system to quickly compare and select optimal pipelines without performing full feature calculations during the search phase, thereby reducing pipeline construction time while maintaining model accuracy.
Solution Approach 2:
The patent introduces meta-features as an intermediary representation that summarizes complex time series data characteristics. Instead of directly comparing raw time series data or performing exhaustive parameter searches, the system uses meta-features as a mediator to efficiently evaluate and compare different lookback window parameters, reducing the computational burden while preserving model performance.
2Reliability
If exhaustive search of lookback window parameters is performed to find optimal parameters, then model performance is improved, but the search process becomes computationally expensive and time-consuming
Solution Approach 1:
The patent extracts essential characteristics of time series data into meta-features (such as statistical properties, trend indicators, and seasonality metrics). By taking out only the most relevant features rather than performing exhaustive analysis of all possible parameters, the system maintains model performance while significantly improving search efficiency.
Solution Approach 2:
The patent transforms the complex parameter search problem into a meta-feature space where optimization is more efficient. By changing the representation from raw parameter values to meta-feature profiles, the system can more effectively identify optimal lookback window parameters without exhaustive search, balancing model performance and search productivity.
3Adaptability or versatility
If the search space for lookback window parameters is expanded to cover more possibilities, then the chance of finding optimal parameters increases, but the complexity of the search process increases
Solution Approach 1:
The patent transitions the search problem from the original parameter space to a meta-feature space, effectively changing dimensions of the search problem. This dimensional transformation allows the system to explore a broader range of lookback window parameter possibilities while managing complexity through the structured meta-feature representation, which organizes diverse parameters into comparable categories.
Data Source
AI summary
Predefined pipelines may be created with predefined meta-features. Time series data may be segmented using lookback window parameters. Meta-features may be determined for windowed data. Those of the predefined pipelines having a maximum amount of matching predefined meta-features may be determined. Those of the lookback window parameters that result in the windowed data having the meta-features most similar to the meta-features of one or more of the plurality of predefined pipelines may be identified.