One-Pass Timeseries Forecasting with Proxy Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning (ML) approaches for timeseries forecasting face challenges in automating model selection and hyperparameter tuning, particularly for datasets outside their tuned scope, leading to inefficient and error-prone processes, and lack of effective comparison methods for different algorithms, resulting in suboptimal performance and high runtime costs.
Innovation Solution
A one-pass ML pipeline that includes an advanced timeseries preprocessor, proxy ML models, and an adaptive cross-validation strategy to automatically select the best algorithm and tune hyperparameters based on temporal statistics, enabling efficient and accurate forecasting without requiring predefined forecast horizons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If exhaustive evaluation of all algorithm and hyperparameter combinations is performed, then model selection accuracy is improved, but runtime cost increases exponentially
Solution Approach 1:
The patent creates proxy models that are simplified versions of the full ML pipeline. These proxy models replicate the essential behavior of complex algorithms and hyperparameter combinations without requiring full training and evaluation. By evaluating proxy models instead of full models, the system achieves accurate model selection with dramatically reduced computational cost, resolving the contradiction between selection accuracy and runtime expense.
Solution Approach 2:
The system performs preliminary analysis of timeseries data to compute temporal statistics and characteristics before model selection. This preliminary action includes calculating autocorrelation functions, detecting seasonality, and assessing stationarity. These pre-computed statistics serve as inputs for proxy model evaluation, enabling accurate algorithm selection without exhaustively testing all hyperparameter combinations, thus reducing runtime while maintaining selection quality.
2Adaptability or versatility
If multiple ML algorithms are evaluated for timeseries forecasting, then algorithm versatility is improved, but system complexity increases
Solution Approach 1:
The patent segments the model selection process into distinct phases: data preprocessing, temporal statistics computation, proxy model evaluation, and final algorithm selection. Each phase handles specific aspects of the complexity independently. The preprocessing module handles data cleaning and transformation, while the evaluation module uses simplified proxy models to compare algorithms. This segmentation reduces overall system complexity by dividing the complex task into manageable, independent components that can be executed systematically.
Solution Approach 2:
The patent introduces proxy models as intermediary objects between the raw timeseries data and the final algorithm selection. These proxy models simplify the complex ML algorithms into easier-to-evaluate representations that capture essential algorithmic behavior. By using proxies as intermediaries, the system can compare multiple algorithms without directly implementing and training all of them, thus reducing system complexity while maintaining versatility in algorithm evaluation.
3Ease of manufacture
If forecast horizon is specified in advance, then model training is simplified, but performance degrades for unexpected forecast periods
Solution Approach 1:
The patent implements a dynamic model selection process where the evaluation criteria and proxy models automatically adapt to the characteristics of the input timeseries data. Rather than using a fixed forecast horizon, the system computes temporal statistics from the data itself and uses these to dynamically determine appropriate evaluation parameters. This dynamic adaptation allows the same pipeline to handle various forecast horizons and data characteristics without requiring manual reconfiguration, resolving the contradiction between training simplicity and forecast adaptability.
Data Source
AI summary
Herein are timeseries preprocessing, model selection, and hyperparameter tuning techniques for forecasting development based on temporal statistics of a timeseries and a single feed-forward pass through a machine learning (ML) pipeline. In an embodiment, a computer hosts and operates the ML pipeline that automatically measures temporal statistic(s) of a timeseries. ML algorithm selection, cross validation, and hyperparameters tuning is based on the temporal statistics of the timeseries. The result from the ML pipeline is a rigorously trained and production ready ML model that is validated to have increased accuracy for multiple prediction horizons. Based on the temporal statistics, efficiency is achieved by asymmetry of investment of computer resources in the tuning and training of the most promising ML algorithm(s). Compared to other approaches, this ML pipeline produces a more accurate ML model for a given amount of computer resources and consumes fewer computer resources to achieve a given accuracy.