Learning Device Quantile Transformation for Distribution Shift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing learning models face challenges when transitioning from training data to decision-making scenarios, as the distribution of variable values may differ between learning and operation, leading to potential inaccuracies in predictions or evaluations.
Innovation Solution
A learning device and method that calculates estimation target item reference values and acquires learning data including fixed and variable values, training models to output estimated values based on these inputs, with evaluation functions that prioritize accuracy across different value distributions, such as using inter-distribution distance reduction and independence indices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the model is trained using fixed values and variable values from learning data, then the model can make predictions for decision making, but the distribution of variable values may differ between learning and operation, leading to prediction inaccuracies
Solution Approach 1:
The patent transforms the original variable values into new variable values by applying parameter changes. Specifically, it uses quantile transformation to map the original variable values to a standard normal distribution, then transforms them to match the target distribution. This parameter transformation ensures that the model trained on transformed data can generalize better to operational data with different distributions, thereby resolving the contradiction between prediction accuracy and distribution difference.
2Measurement precision
If the model is trained to minimize error between estimated values and actual values, then prediction accuracy improves, but the model may not handle distribution shifts effectively when deployed
Solution Approach 1:
The patent applies preliminary action by transforming the variable values before model training. The quantile transformation and target distribution matching are performed in advance on the training data, so that when the model is trained, it learns from data that already accounts for the distribution characteristics. This preliminary transformation ensures that the model becomes robust to distribution shifts when deployed, while maintaining high estimation accuracy on the transformed scale.
3Adaptability or versatility
If the model uses transformed variable values to match target distribution, then the model generalizes better to operational data, but the transformation process adds complexity to the learning system
Solution Approach 1:
The patent uses copying by creating a copy of the variable values and applying transformations only to this copy, rather than modifying the original data structure or creating complex transformation pipelines. The quantile transformation and target distribution matching are applied to generate new variable values that are then used for training. This copying approach simplifies the overall system complexity while achieving distribution matching, as the transformation can be implemented as a separate, standalone process.
Data Source
AI summary
A learning device calculates an estimation target item reference value according to a fixed value of each estimation target object. The learning device acquires learning data that includes the fixed value of each estimation target object, a variable item value, and an estimation target item value according to the fixed value and the variable item value. The learning device trains, using the learning data and an evaluation function, a model that outputs an estimated value of the estimation target item value in response to input of the fixed value of each estimation target object and the variable item value.