Time-Varying Feature Engineering for Cloud Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine learning forecasting techniques for cloud services fail to detect complex time-based relationships within input data, relying on static features that do not fully utilize metadata to improve forecasting accuracy.
Innovation Solution
The method involves feature engineering to generate time-varying features by combining time-series data with metadata, which describes relationships between data points, and incorporating exogenous data to enhance the input for machine learning forecasting models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional static features are used in machine learning forecasting, then the forecasting model is simple to implement, but the forecasting accuracy is insufficient due to failure to detect complex time-based relationships
Solution Approach 1:
The patent transforms static features into dynamic time-varying features that evolve over time. The system generates features at multiple time steps (t-1, t-2, ..., t-n) that capture temporal relationships, allowing the forecasting model to adapt to changing patterns in the data while maintaining structured feature engineering through automated template application.
Solution Approach 2:
The patent adds a temporal dimension to traditional static features by creating time-varying features that incorporate historical data at multiple time steps. This transforms univariate static features into multivariate time-series features, enabling the model to detect complex time-based relationships without manually engineering each temporal relationship.
2Measurement precision
If multivariate time-varying features are generated from time-series data and metadata, then the forecasting accuracy improves by capturing temporal relationships, but the data processing complexity increases
Solution Approach 1:
The system automatically generates metadata that describes relationships between time-series data points, eliminating the need for manual relationship specification. The metadata generation process self-services by extracting relationship information directly from the time-series data, and the feature engineering templates automatically apply appropriate transformations based on detected relationships.
Solution Approach 2:
The patent creates a universal feature engineering framework that handles multiple relationship types (trend, seasonality, autocorrelation) through standardized templates. These templates can be applied across different time-series datasets regardless of their specific characteristics, reducing the complexity of detecting and measuring relationships by providing a unified approach.
3Measurement precision
If exogenous data is incorporated into the forecasting model, then the model captures external influences improving accuracy, but the model complexity and computational requirements increase
Solution Approach 1:
The patent merges endogenous time-series features with exogenous data features into a unified feature set. Both types of features are processed through the same feature engineering templates and relationship detection mechanisms, allowing the model to capture external influences without requiring separate processing pipelines or increased model complexity.
Data Source
AI summary
The present embodiments relate to using feature engineering to generate time-varying features via metadata. A first exemplary embodiment provides a method for performing feature engineering to generate time-varying features. The method can include receiving a first value and a second value of the time-series data. The method can further include receiving metadata that describes a relationship between the first value and the second value. The method can further include detecting the relationship between the first value and the second value based on the metadata. The method can further include generating, a time-varying feature from a combination of the first value and the second value based on the relationship detected from the metadata. The method can further include generating, by implementing the machine learning forecasting model, a forecasted value for the time-series data based on the time-varying feature.