AI Model Training for Baseline Shift and Moving Classifications
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial intelligence models relying on time series data are susceptible to baseline shifts, leading to inaccurate predictions, especially in scenarios with short outlier events, as they are trained on historic data over long periods, and existing methods to address these shifts either eliminate valuable data or introduce systematic bias.
Innovation Solution
The system generates two classes of training data based on a known baseline shift, allowing for the training of multiple models or a single model in one step, using a uniform feature input and time criteria to parse out model signals from anomalous time periods, preserving long-term trend accuracy while eliminating bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the model is trained on historic data over long periods of time, then the model achieves accuracy for general long-term trends, but the model misses short outlier events and has lengthy training lags upon baseline shifts
Solution Approach 1:
The patent segments the training data into multiple classes based on baseline shifts. Each class corresponds to a specific time period or baseline condition, allowing the model to process and learn from different temporal segments separately. This segmentation enables the model to maintain accuracy for long-term trends while reducing training lag when baseline shifts occur, as each segment can be processed independently without requiring complete retraining on all historic data.
2Measurement precision
If the system determines expiration dates for training data to account for outliers, then short-term accuracy improves, but valuable data that indicates long-term trends is eliminated
Solution Approach 1:
Instead of eliminating data based on expiration dates, the patent segments training data into multiple classes based on baseline shifts. Each class preserves data from specific time periods, allowing the system to maintain both short-term and long-term information. This segmentation approach enables the model to access relevant data for different time horizons without discarding valuable long-term trend information.
Solution Approach 2:
The patent applies local quality by treating different time periods and baseline conditions differently through class-specific processing. Each class of training data is processed according to its specific temporal characteristics and baseline conditions, allowing the model to optimize for both short-term precision and long-term trend accuracy without uniform treatment that would discard valuable information.
3Measurement precision
If the system tags data as outlier data to account for baseline shifts, then short-term prediction improves, but systematic bias is introduced into the modeling process
Solution Approach 1:
The patent segments training data into multiple classes based on baseline shifts rather than tagging individual outlier records. This segmentation approach classifies data by temporal and baseline characteristics, allowing the model to learn patterns from different periods without introducing systematic bias. By processing data through class-specific models or weighted averages, the system achieves short-term accuracy without the bias introduced by selective tagging.
4Reliability
If the model processes all training data uniformly, then long-term trend accuracy is maintained, but the model fails to adapt quickly to baseline shifts
Solution Approach 1:
The patent segments training data into multiple classes based on baseline shifts, enabling the model to process different time periods and baseline conditions separately. This segmentation allows the model to maintain long-term trend accuracy through comprehensive data processing while improving adaptability to baseline shifts by processing recent data through specialized class-specific pathways that prioritize recent patterns.
Solution Approach 2:
The patent introduces dynamic processing by adjusting the weightings and processing methods for different data classes based on their temporal proximity and baseline characteristics. The system dynamically adjusts which classes receive higher processing priority and how their data is weighted in the final model output, enabling quick adaptation to baseline shifts while maintaining long-term accuracy through balanced processing of historic data.
Data Source
AI summary
The system generates multiple classes of training data depending on their baseline shift. The system then trains a model or ensemble of models leveraging the entire dataset without the need for training them separately. A set of uniform input feature are generated for a user based on profile of the data (e.g., based on a user's auto loan terms and/or credit history). The system then engineers additional characteristic features based on the time criteria (e.g., based on a time period overlap with the time period of the classes and what class the profile data originated from). The system uses this time criteria (e.g., whether profile data originated from a given class and the amount of overlap of the data with a class) to determine how the feature input is processed through the models.