Time Series Model Reduction via Dimensional Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models trained on separate portions of large-scale time series data require significant computational resources, are sensitive to missing data and noise, and often result in overfitting, leading to inaccurate predictions.
Innovation Solution
The system employs model reduction techniques by encoding time series data into a lower dimension space, generating prediction models in this space, and then decoding the outputs back to the original dimension space, using techniques like factorization and graph-regularized alternating least squares to reduce training time and improve robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple models are trained separately on portions of the data pool, then each model can be trained on specific location data, but the computational resources required increase significantly and training time increases
Solution Approach 1:
The patent segments the data pool into location-specific portions while training a single unified model. The model is trained on the entire data pool with location identifiers, allowing the model to learn location-specific patterns without requiring separate models for each location. This segmentation approach reduces computational resources while maintaining location-specific adaptability.
Solution Approach 2:
The patent creates a universal model that can handle multiple locations simultaneously. The single model is designed to process data from different locations by incorporating location identifiers, making it multi-functional across different geographic regions. This eliminates the need for multiple separate models while maintaining the ability to provide location-specific predictions.
2Adaptability or versatility
If models are trained separately on portions of the data pool, then each model focuses on specific location data, but the models become sensitive to missing data and noise in their limited training portions
Solution Approach 1:
The patent merges multiple location-specific training tasks into a single unified model training process. By training one model on the entire data pool rather than separate models on portions, the system benefits from the combined data volume, which reduces sensitivity to missing data and noise in any single location. The unified model learns robust patterns across all locations simultaneously.
Solution Approach 2:
The patent uses location identifiers to create virtual copies of location-specific information within the unified model. Instead of training separate models, the single model stores and processes location-specific patterns through the identifier system, allowing it to replicate location-specific predictions while benefiting from the entire data pool's robustness.
3Adaptability or versatility
If models are trained separately on portions of the data pool, then each model is trained on limited data, but this results in overfitting to the specific portion and inaccurate predictions
Solution Approach 1:
The patent adds a location identifier dimension to the data structure, transforming the problem from training multiple separate models to training a single model that operates in an enhanced feature space. This dimensional addition allows the model to distinguish between locations while maintaining a unified training process, preventing overfitting by distributing learning across the entire data pool rather than isolating it to portions.
4Use of energy by moving object
If the entire data pool is used for training a single model, then computational resources are reduced, but the model cannot provide location-specific predictions
Solution Approach 1:
The patent introduces location identifiers as intermediary features that bridge the gap between unified training and location-specific predictions. These identifiers act as mediators that allow the single model to distinguish and process location-specific information while maintaining a unified training architecture. The identifiers enable the model to provide location-specific predictions without requiring separate models for each location.
Data Source
AI summary
Provided are systems for generating a machine learning model and a prediction based on encoded time series data using model reduction techniques that include a processor to receive a training dataset of a plurality of data instances, wherein each data instance includes a time series of data points, perform an encoding operation on the training dataset to provide an encoded dataset having a lower dimension space than a dimension space of the training dataset, generate one or more prediction models based on the encoded dataset, determine an output of the one or more prediction models in the lower dimension space based on an input provided to the one or more prediction models, and perform a decoding operation on the output to project the output from the lower dimension space to the dimension space of the training dataset. Methods and computer program products are also provided.